Merge pull request #2006 from roboflow/feature/xyxy_to_mask

Feature/xyxy to mask

.github/workflows/publish-docs.yml

@@ -34,7 +34,7 @@ jobs:
           fetch-depth: 0
 
       - name: 🐍 Install uv and set Python ${{ matrix.python-version }}
-        uses: astral-sh/setup-uv@bd01e18f51369d5a26f1651c3cb451d3417e3bba # v6.3.1
+        uses: astral-sh/setup-uv@e92bafb6253dcd438e0484186d7669ea7a8ca1cc # v6.4.3
         with:
           python-version: ${{ matrix.python-version }}
           activate-environment: true

.github/workflows/publish-pre-release.yml

@@ -29,7 +29,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: 🐍 Install uv and set Python version ${{ matrix.python-version }}
-        uses: astral-sh/setup-uv@bd01e18f51369d5a26f1651c3cb451d3417e3bba # v6.3.1
+        uses: astral-sh/setup-uv@e92bafb6253dcd438e0484186d7669ea7a8ca1cc # v6.4.3
         with:
           python-version: ${{ matrix.python-version }}
           activate-environment: true

.github/workflows/publish-release.yml

@@ -27,7 +27,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: 🐍 Install uv and set Python version ${{ matrix.python-version }}
-        uses: astral-sh/setup-uv@bd01e18f51369d5a26f1651c3cb451d3417e3bba # v6.3.1
+        uses: astral-sh/setup-uv@e92bafb6253dcd438e0484186d7669ea7a8ca1cc # v6.4.3
         with:
           python-version: ${{ matrix.python-version }}
           activate-environment: true

.github/workflows/publish-testpypi.yml

@@ -24,7 +24,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: 🐍 Install uv and set Python version ${{ matrix.python-version }}
-        uses: astral-sh/setup-uv@bd01e18f51369d5a26f1651c3cb451d3417e3bba # v6.3.1
+        uses: astral-sh/setup-uv@e92bafb6253dcd438e0484186d7669ea7a8ca1cc # v6.4.3
         with:
           python-version: ${{ matrix.python-version }}
           activate-environment: true

.github/workflows/test-doc.yml

@@ -24,7 +24,7 @@ jobs:
           fetch-depth: 0
 
       - name: 🐍 Install uv and set Python ${{ matrix.python-version }}
-        uses: astral-sh/setup-uv@bd01e18f51369d5a26f1651c3cb451d3417e3bba # v6.3.1
+        uses: astral-sh/setup-uv@e92bafb6253dcd438e0484186d7669ea7a8ca1cc # v6.4.3
         with:
           python-version: ${{ matrix.python-version }}
           activate-environment: true

.github/workflows/uv-test.yml

@@ -19,7 +19,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: 🐍 Install uv and set Python version ${{ matrix.python-version }}
-        uses: astral-sh/setup-uv@bd01e18f51369d5a26f1651c3cb451d3417e3bba # v6.3.1
+        uses: astral-sh/setup-uv@e92bafb6253dcd438e0484186d7669ea7a8ca1cc # v6.4.3
         with:
           python-version: ${{ matrix.python-version }}
           activate-environment: true

.pre-commit-config.yaml

@@ -32,7 +32,7 @@ repos:
           additional_dependencies: ["bandit[toml]"]
 
   -   repo: https://github.com/astral-sh/ruff-pre-commit
-      rev: v0.12.3
+      rev: v0.12.5
       hooks:
       -   id: ruff
           args: [--fix, --exit-non-zero-on-fix]

docs/detection/utils/converters.md

@@ -58,3 +58,9 @@ status: new
 </div>
 
 :::supervision.detection.utils.converters.polygon_to_xyxy
+
+<div class="md-typeset">
+  <h2><a href="#supervision.detection.utils.converters.xyxy_to_mask">xyxy_to_mask</a></h2>
+</div>
+
+:::supervision.detection.utils.converters.xyxy_to_mask

docs/detection/utils/vlms.md

@@ -0,0 +1,18 @@
+---
+comments: true
+status: new
+---
+
+# VLMs Utils
+
+<div class="md-typeset">
+  <h2><a href="#supervision.detection.utils.vlms.edit_distance">edit_distance</a></h2>
+</div>
+
+:::supervision.detection.utils.vlms.edit_distance
+
+<div class="md-typeset">
+  <h2><a href="#supervision.detection.utils.vlms.fuzzy_match_index">fuzzy_match_index</a></h2>
+</div>
+
+:::supervision.detection.utils.vlms.fuzzy_match_index

docs/how_to/process_datasets.md

@@ -331,12 +331,9 @@ for i in range(16):
     annotated_image = label_annotator.annotate(annotated_image, annotations, labels)
     annotated_images.append(annotated_image)
 
-grid = sv.create_tiles(
+sv.plot_images_grid(
     annotated_images,
     grid_size=(4, 4),
-    single_tile_size=(400, 400),
-    tile_padding_color=sv.Color.WHITE,
-    tile_margin_color=sv.Color.WHITE
 )
 ```
 

docs/how_to/track_objects.md

@@ -345,7 +345,7 @@ Supervision is versatile and compatible with various models. Check this [link](/
 
 We will define a `callback` function, which will process each frame of the video by obtaining model predictions and then annotating the frame based on these predictions.
 
-Let's immediately visualize the results with our [`EdgeAnnotator`](/latest/keypoint/annotators/#supervision.keypoint.annotators.EdgeAnnotator) and [`VertexAnnotator`](https://supervision.roboflow.com/latest/keypoint/annotators/#supervision.keypoint.annotators.VertexAnnotator).
+Let's immediately visualize the results with our [`EdgeAnnotator`](/latest/keypoint/annotators/#supervision.key_points.annotators.EdgeAnnotator) and [`VertexAnnotator`](https://supervision.roboflow.com/latest/keypoint/annotators/#supervision.key_points.annotators.VertexAnnotator).
 
 === "Ultralytics"
 
@@ -408,7 +408,7 @@ Let's immediately visualize the results with our [`EdgeAnnotator`](/latest/keypo
 
 ### Convert to Detections
 
-Keypoint tracking is currently supported via the conversion of `KeyPoints` to `Detections`. This is achieved with the [`KeyPoints.as_detections()`](/latest/keypoint/core/#supervision.keypoint.core.KeyPoints.as_detections) function.
+Keypoint tracking is currently supported via the conversion of `KeyPoints` to `Detections`. This is achieved with the [`KeyPoints.as_detections()`](/latest/keypoint/core/#supervision.key_points.core.KeyPoints.as_detections) function.
 
 Let's convert to detections and visualize the results with our [`BoxAnnotator`](/latest/detection/annotators/#supervision.annotators.core.BoxAnnotator).
 

docs/keypoint/annotators.md

@@ -78,19 +78,19 @@ comments: true
     </div>
 
 <div class="md-typeset">
-  <h2><a href="#supervision.keypoint.annotators.VertexAnnotator">VertexAnnotator</a></h2>
+  <h2><a href="#supervision.key_points.annotators.VertexAnnotator">VertexAnnotator</a></h2>
 </div>
 
-:::supervision.keypoint.annotators.VertexAnnotator
+:::supervision.key_points.annotators.VertexAnnotator
 
 <div class="md-typeset">
-  <h2><a href="#supervision.keypoint.annotators.EdgeAnnotator">EdgeAnnotator</a></h2>
+  <h2><a href="#supervision.key_points.annotators.EdgeAnnotator">EdgeAnnotator</a></h2>
 </div>
 
-:::supervision.keypoint.annotators.EdgeAnnotator
+:::supervision.key_points.annotators.EdgeAnnotator
 
 <div class="md-typeset">
-  <h2><a href="#supervision.keypoint.annotators.VertexLabelAnnotator">VertexLabelAnnotator</a></h2>
+  <h2><a href="#supervision.key_points.annotators.VertexLabelAnnotator">VertexLabelAnnotator</a></h2>
 </div>
 
-:::supervision.keypoint.annotators.VertexLabelAnnotator
+:::supervision.key_points.annotators.VertexLabelAnnotator

docs/keypoint/core.md

@@ -5,4 +5,4 @@ status: new
 
 # Keypoint Detection
 
-:::supervision.keypoint.core.KeyPoints
+:::supervision.key_points.core.KeyPoints

docs/utils/image.md

@@ -1,5 +1,6 @@
 ---
 comments: true
+status: new
 ---
 
 # Image Utils
@@ -29,10 +30,16 @@ comments: true
 :::supervision.utils.image.letterbox_image
 
 <div class="md-typeset">
-    <h2><a href="#supervision.utils.image.overlay_image">overlay_image</a></h2>
+    <h2><a href="#supervision.utils.image.tint_image">tint_image</a></h2>
 </div>
 
-:::supervision.utils.image.overlay_image
+:::supervision.utils.image.tint_image
+
+<div class="md-typeset">
+    <h2><a href="#supervision.utils.image.grayscale_image">grayscale_image</a></h2>
+</div>
+
+:::supervision.utils.image.grayscale_image
 
 <div class="md-typeset">
     <h2><a href="#supervision.utils.image.ImageSink">ImageSink</a></h2>

examples/time_in_zone/requirements.txt

@@ -1,4 +1,6 @@
 supervision
 ultralytics
 inference
-pytube
+# https://github.com/pytube/pytube/issues/2044
+# pytube
+pytubefix

examples/time_in_zone/scripts/download_from_youtube.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 import argparse
 import os
 
-from pytube import YouTube
+from pytubefix import YouTube
 
 
 def main(url: str, output_path: str | None, file_name: str | None) -> None:

mkdocs.yml

@@ -26,7 +26,7 @@ extra:
 
 extra_css:
   - stylesheets/extra.css
-  - stylesheets/cookbooks-card.css
+  - stylesheets/cookbooks_card.css
 
 nav:
   - Home: index.md
@@ -47,6 +47,7 @@ nav:
           - Boxes: detection/utils/boxes.md
           - Masks: detection/utils/masks.md
           - Polygons: detection/utils/polygons.md
+          - VLMs: detection/utils/vlms.md
       - Keypoint Detection:
           - Core: keypoint/core.md
           - Annotators: keypoint/annotators.md

pyproject.toml

@@ -2,7 +2,7 @@
 name = "supervision"
 description = "A set of easy-to-use utils that will come in handy in any Computer Vision project"
 license = { text = "MIT" }
-version = "0.26.1"
+version = "0.27.0rc2"
 readme = "README.md"
 requires-python = ">=3.9"
 authors = [

supervision/__init__.py

@@ -65,6 +65,7 @@ from supervision.detection.utils.converters import (
     polygon_to_xyxy,
     xcycwh_to_xyxy,
     xywh_to_xyxy,
+    xyxy_to_mask,
     xyxy_to_polygons,
     xyxy_to_xcycarh,
     xyxy_to_xywh,
@@ -92,6 +93,7 @@ from supervision.detection.utils.polygons import (
     approximate_polygon,
     filter_polygons_by_area,
 )
+from supervision.detection.utils.vlms import edit_distance, fuzzy_match_index
 from supervision.detection.vlm import LMM, VLM
 from supervision.draw.color import Color, ColorPalette
 from supervision.draw.utils import (
@@ -107,24 +109,25 @@ from supervision.draw.utils import (
 )
 from supervision.geometry.core import Point, Position, Rect
 from supervision.geometry.utils import get_polygon_center
-from supervision.keypoint.annotators import (
+from supervision.key_points.annotators import (
     EdgeAnnotator,
     VertexAnnotator,
     VertexLabelAnnotator,
 )
-from supervision.keypoint.core import KeyPoints
+from supervision.key_points.core import KeyPoints
 from supervision.metrics.detection import ConfusionMatrix, MeanAveragePrecision
 from supervision.tracker.byte_tracker.core import ByteTrack
 from supervision.utils.conversion import cv2_to_pillow, pillow_to_cv2
 from supervision.utils.file import list_files_with_extensions
 from supervision.utils.image import (
     ImageSink,
-    create_tiles,
     crop_image,
+    grayscale_image,
     letterbox_image,
     overlay_image,
     resize_image,
     scale_image,
+    tint_image,
 )
 from supervision.utils.notebook import plot_image, plot_images_grid
 from supervision.utils.video import (
@@ -205,7 +208,6 @@ __all__ = [
     "clip_boxes",
     "contains_holes",
     "contains_multiple_segments",
-    "create_tiles",
     "crop_image",
     "cv2_to_pillow",
     "draw_filled_polygon",
@@ -215,10 +217,13 @@ __all__ = [
     "draw_polygon",
     "draw_rectangle",
     "draw_text",
+    "edit_distance",
     "filter_polygons_by_area",
+    "fuzzy_match_index",
     "get_coco_class_index_mapping",
     "get_polygon_center",
     "get_video_frames_generator",
+    "grayscale_image",
     "letterbox_image",
     "list_files_with_extensions",
     "mask_iou_batch",
@@ -242,8 +247,10 @@ __all__ = [
     "rle_to_mask",
     "scale_boxes",
     "scale_image",
+    "tint_image",
     "xcycwh_to_xyxy",
     "xywh_to_xyxy",
+    "xyxy_to_mask",
     "xyxy_to_polygons",
     "xyxy_to_xcycarh",
     "xyxy_to_xywh",

supervision/annotators/base.py

@@ -1,19 +1,7 @@
 from abc import ABC, abstractmethod
-from typing import TypeVar
-
-import numpy as np
-from PIL import Image
 
 from supervision.detection.core import Detections
-
-ImageType = TypeVar("ImageType", np.ndarray, Image.Image)
-"""
-An image of type `np.ndarray` or `PIL.Image.Image`.
-
-Unlike a `Union`, ensures the type remains consistent. If a function
-takes an `ImageType` argument and returns an `ImageType`, when you
-pass an `np.ndarray`, you will get an `np.ndarray` back.
-"""
+from supervision.draw.base import ImageType
 
 
 class BaseAnnotator(ABC):

supervision/annotators/core.py

@@ -7,8 +7,9 @@ import cv2
 import numpy as np
 import numpy.typing as npt
 from PIL import Image, ImageDraw, ImageFont
+from scipy.interpolate import splev, splprep
 
-from supervision.annotators.base import BaseAnnotator, ImageType
+from supervision.annotators.base import BaseAnnotator
 from supervision.annotators.utils import (
     PENDING_TRACK_ID,
     ColorLookup,
@@ -28,12 +29,13 @@ from supervision.detection.utils.converters import (
     polygon_to_mask,
     xyxy_to_polygons,
 )
+from supervision.draw.base import ImageType
 from supervision.draw.color import Color, ColorPalette
 from supervision.draw.utils import draw_polygon, draw_rounded_rectangle, draw_text
 from supervision.geometry.core import Point, Position, Rect
 from supervision.utils.conversion import (
-    ensure_cv2_image_for_annotation,
-    ensure_pil_image_for_annotation,
+    ensure_cv2_image_for_class_method,
+    ensure_pil_image_for_class_method,
 )
 from supervision.utils.image import (
     crop_image,
@@ -51,25 +53,28 @@ class _BaseLabelAnnotator(BaseAnnotator):
 
     Attributes:
         color (Union[Color, ColorPalette]): The color to use for the label background.
+        color_lookup (ColorLookup): The method used to determine the color of the label.
         text_color (Union[Color, ColorPalette]): The color to use for the label text.
         text_padding (int): The padding around the label text, in pixels.
         text_anchor (Position): The position of the text relative to the detection
-                                bounding box.
-        color_lookup (ColorLookup): The method used to determine the color of the label.
+            bounding box.
+        text_offset (Tuple[int, int]): A tuple of 2D coordinates `(x, y)` to
+            offset the text position from the anchor point, in pixels.
         border_radius (int): The radius of the label background corners, in pixels.
         smart_position (bool): Whether to intelligently adjust the label position to
-                                avoid overlapping with other elements.
+            avoid overlapping with other elements.
         max_line_length (Optional[int]): Maximum number of characters per line before
-                                wrapping the text. None means no wrapping.
+            wrapping the text. None means no wrapping.
     """
 
     def __init__(
         self,
         color: Color | ColorPalette = ColorPalette.DEFAULT,
+        color_lookup: ColorLookup = ColorLookup.CLASS,
         text_color: Color | ColorPalette = Color.WHITE,
         text_padding: int = 10,
         text_position: Position = Position.TOP_LEFT,
-        color_lookup: ColorLookup = ColorLookup.CLASS,
+        text_offset: tuple[int, int] = (0, 0),
         border_radius: int = 0,
         smart_position: bool = False,
         max_line_length: int | None = None,
@@ -79,27 +84,29 @@ class _BaseLabelAnnotator(BaseAnnotator):
 
         Args:
             color (Union[Color, ColorPalette], optional): The color to use for the label
-                                background.
+                background.
+            color_lookup (ColorLookup, optional): The method used to determine the color
+                of the label
             text_color (Union[Color, ColorPalette], optional): The color to use for the
-                                label text.
+                label text.
             text_padding (int, optional): The padding around the label text, in pixels.
             text_position (Position, optional): The position of the text relative to the
-                                detection bounding box.
-            color_lookup (ColorLookup, optional): The method used to determine the color
-                                of the label
+                detection bounding box.
+            text_offset (Tuple[int, int], optional): A tuple of 2D coordinates
+                `(x, y)` to offset the text position from the anchor point, in pixels.
             border_radius (int, optional): The radius of the label background corners,
-                                in pixels.
+                in pixels.
             smart_position (bool, optional): Whether to intelligently adjust the label
-                                position to avoid overlapping with other elements.
+                position to avoid overlapping with other elements.
             max_line_length (Optional[int], optional): Maximum number of characters per
-                                line before wrapping the text. None means no wrapping.
-
+                line before wrapping the text. None means no wrapping.
         """
         self.color: Color | ColorPalette = color
+        self.color_lookup: ColorLookup = color_lookup
         self.text_color: Color | ColorPalette = text_color
         self.text_padding: int = text_padding
         self.text_anchor: Position = text_position
-        self.color_lookup: ColorLookup = color_lookup
+        self.text_offset: tuple[int, int] = text_offset
         self.border_radius: int = border_radius
         self.smart_position = smart_position
         self.max_line_length: int | None = max_line_length
@@ -171,7 +178,7 @@ class BoxAnnotator(BaseAnnotator):
         self.thickness: int = thickness
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -254,7 +261,7 @@ class OrientedBoxAnnotator(BaseAnnotator):
         self.thickness: int = thickness
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -343,7 +350,7 @@ class MaskAnnotator(BaseAnnotator):
         self.opacity = opacity
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -433,7 +440,7 @@ class PolygonAnnotator(BaseAnnotator):
         self.thickness: int = thickness
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -520,7 +527,7 @@ class ColorAnnotator(BaseAnnotator):
         self.color_lookup: ColorLookup = color_lookup
         self.opacity = opacity
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -616,7 +623,7 @@ class HaloAnnotator(BaseAnnotator):
         self.color_lookup: ColorLookup = color_lookup
         self.kernel_size: int = kernel_size
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -716,7 +723,7 @@ class EllipseAnnotator(BaseAnnotator):
         self.end_angle: int = end_angle
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -808,7 +815,7 @@ class BoxCornerAnnotator(BaseAnnotator):
         self.corner_length: int = corner_length
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -897,7 +904,7 @@ class CircleAnnotator(BaseAnnotator):
         self.thickness: int = thickness
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -996,7 +1003,7 @@ class DotAnnotator(BaseAnnotator):
         self.outline_thickness = outline_thickness
         self.outline_color: Color | ColorPalette = outline_color
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -1076,37 +1083,100 @@ class LabelAnnotator(_BaseLabelAnnotator):
     def __init__(
         self,
         color: Color | ColorPalette = ColorPalette.DEFAULT,
+        color_lookup: ColorLookup = ColorLookup.CLASS,
         text_color: Color | ColorPalette = Color.WHITE,
         text_scale: float = 0.5,
         text_thickness: int = 1,
         text_padding: int = 10,
         text_position: Position = Position.TOP_LEFT,
-        color_lookup: ColorLookup = ColorLookup.CLASS,
+        text_offset: tuple[int, int] = (0, 0),
         border_radius: int = 0,
         smart_position: bool = False,
         max_line_length: int | None = None,
     ):
+        """
+        Args:
+            color (Union[Color, ColorPalette]): The color or color palette to use for
+                annotating the text background.
+            color_lookup (ColorLookup): Strategy for mapping colors to annotations.
+                Options are `INDEX`, `CLASS`, `TRACK`.
+            text_color (Union[Color, ColorPalette]): The color or color palette to use
+                for the text.
+            text_scale (float): Font scale for the text.
+            text_thickness (int): Thickness of the text characters.
+            text_padding (int): Padding around the text within its background box.
+            text_position (Position): Position of the text relative to the detection.
+                Possible values are defined in the `Position` enum.
+            text_offset (Tuple[int, int]): A tuple of 2D coordinates `(x, y)` to
+                offset the text position from the anchor point, in pixels.
+            border_radius (int): The radius to apply round edges. If the selected
+                value is higher than the lower dimension, width or height, is clipped.
+            smart_position (bool): Spread out the labels to avoid overlapping.
+            max_line_length (Optional[int]): Maximum number of characters per line
+                before wrapping the text. None means no wrapping.
+        """
         self.text_scale: float = text_scale
         self.text_thickness: int = text_thickness
         super().__init__(
             color=color,
+            color_lookup=color_lookup,
             text_color=text_color,
             text_padding=text_padding,
             text_position=text_position,
-            color_lookup=color_lookup,
+            text_offset=text_offset,
             border_radius=border_radius,
             smart_position=smart_position,
             max_line_length=max_line_length,
         )
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
-        scene: ImageType,  # Ensure scene is initially a NumPy array here
+        scene: ImageType,
         detections: Detections,
         labels: list[str] | None = None,
         custom_color_lookup: np.ndarray | None = None,
     ) -> np.ndarray:
+        """
+        Annotates the given scene with labels based on the provided detections.
+
+        Args:
+            scene (ImageType): The image where labels will be drawn.
+                `ImageType` is a flexible type, accepting either `numpy.ndarray`
+                or `PIL.Image.Image`.
+            detections (Detections): Object detections to annotate.
+            labels (Optional[List[str]]): Custom labels for each detection.
+            custom_color_lookup (Optional[np.ndarray]): Custom color lookup array.
+                Allows to override the default color mapping strategy.
+
+        Returns:
+            The annotated image, matching the type of `scene` (`numpy.ndarray`
+                or `PIL.Image.Image`)
+
+        Example:
+            ```python
+            import supervision as sv
+
+            image = ...
+            detections = sv.Detections(...)
+
+            labels = [
+                f"{class_name} {confidence:.2f}"
+                for class_name, confidence
+                in zip(detections['class_name'], detections.confidence)
+            ]
+
+            label_annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER)
+            annotated_frame = label_annotator.annotate(
+                scene=image.copy(),
+                detections=detections,
+                labels=labels
+            )
+            ```
+
+        ![label-annotator-example](https://media.roboflow.com/
+        supervision-annotator-examples/label-annotator-example-purple.png)
+        """
         assert isinstance(scene, np.ndarray)
         validate_labels(labels, detections)
 
@@ -1144,7 +1214,12 @@ class LabelAnnotator(_BaseLabelAnnotator):
             anchor=self.text_anchor
         ).astype(int)
 
-        for label, center_coords in zip(labels, anchors_coordinates):
+        for label, center_coordinates in zip(labels, anchors_coordinates):
+            center_coordinates = (
+                center_coordinates[0] + self.text_offset[0],
+                center_coordinates[1] + self.text_offset[1],
+            )
+
             wrapped_lines = wrap_text(label, self.max_line_length)
             line_heights = []
             line_widths = []
@@ -1170,7 +1245,7 @@ class LabelAnnotator(_BaseLabelAnnotator):
             height_padded = total_height + 2 * self.text_padding
 
             text_background_xyxy = resolve_text_background_xyxy(
-                center_coordinates=tuple(center_coords),
+                center_coordinates=center_coordinates,
                 text_wh=(width_padded, height_padded),
                 position=self.text_anchor,
             )
@@ -1317,31 +1392,54 @@ class RichLabelAnnotator(_BaseLabelAnnotator):
     def __init__(
         self,
         color: Color | ColorPalette = ColorPalette.DEFAULT,
+        color_lookup: ColorLookup = ColorLookup.CLASS,
         text_color: Color | ColorPalette = Color.WHITE,
         font_path: str | None = None,
         font_size: int = 10,
         text_padding: int = 10,
         text_position: Position = Position.TOP_LEFT,
-        color_lookup: ColorLookup = ColorLookup.CLASS,
+        text_offset: tuple[int, int] = (0, 0),
         border_radius: int = 0,
         smart_position: bool = False,
         max_line_length: int | None = None,
     ):
+        """
+        Args:
+            color (Union[Color, ColorPalette]): The color or color palette to use for
+                annotating the text background.
+            color_lookup (ColorLookup): Strategy for mapping colors to annotations.
+                Options are `INDEX`, `CLASS`, `TRACK`.
+            text_color (Union[Color, ColorPalette]): The color to use for the text.
+            font_path (Optional[str]): Path to the font file (e.g., ".ttf" or ".otf")
+                to use for rendering text. If `None`, the default PIL font will be used.
+            font_size (int): Font size for the text.
+            text_padding (int): Padding around the text within its background box.
+            text_position (Position): Position of the text relative to the detection.
+                Possible values are defined in the `Position` enum.
+            text_offset (Tuple[int, int]): A tuple of 2D coordinates `(x, y)` to
+                offset the text position from the anchor point, in pixels.
+            border_radius (int): The radius to apply round edges. If the selected
+                value is higher than the lower dimension, width or height, is clipped.
+            smart_position (bool): Spread out the labels to avoid overlapping.
+            max_line_length (Optional[int]): Maximum number of characters per line
+                before wrapping the text. None means no wrapping.
+        """
         self.font_path = font_path
         self.font_size = font_size
         self.font = self._load_font(font_size, font_path)
         super().__init__(
             color=color,
+            color_lookup=color_lookup,
             text_color=text_color,
             text_padding=text_padding,
             text_position=text_position,
-            color_lookup=color_lookup,
+            text_offset=text_offset,
             border_radius=border_radius,
             smart_position=smart_position,
             max_line_length=max_line_length,
         )
 
-    @ensure_pil_image_for_annotation
+    @ensure_pil_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -1349,6 +1447,44 @@ class RichLabelAnnotator(_BaseLabelAnnotator):
         labels: list[str] | None = None,
         custom_color_lookup: np.ndarray | None = None,
     ) -> ImageType:
+        """
+        Annotates the given scene with labels based on the provided
+        detections, with support for Unicode characters.
+
+        Args:
+            scene (ImageType): The image where labels will be drawn.
+                `ImageType` is a flexible type, accepting either `numpy.ndarray`
+                or `PIL.Image.Image`.
+            detections (Detections): Object detections to annotate.
+            labels (Optional[List[str]]): Custom labels for each detection.
+            custom_color_lookup (Optional[np.ndarray]): Custom color lookup array.
+                Allows to override the default color mapping strategy.
+
+        Returns:
+            The annotated image, matching the type of `scene` (`numpy.ndarray`
+                or `PIL.Image.Image`)
+
+        Example:
+            ```python
+            import supervision as sv
+
+            image = ...
+            detections = sv.Detections(...)
+
+            labels = [
+                f"{class_name} {confidence:.2f}"
+                for class_name, confidence
+                in zip(detections['class_name'], detections.confidence)
+            ]
+
+            rich_label_annotator = sv.RichLabelAnnotator(font_path="path/to/font.ttf")
+            annotated_frame = label_annotator.annotate(
+                scene=image.copy(),
+                detections=detections,
+                labels=labels
+            )
+            ```
+        """
         assert isinstance(scene, Image.Image)
         validate_labels(labels, detections)
 
@@ -1386,7 +1522,12 @@ class RichLabelAnnotator(_BaseLabelAnnotator):
             anchor=self.text_anchor
         ).astype(int)
 
-        for label, center_coords in zip(labels, anchor_coordinates):
+        for label, center_coordinates in zip(labels, anchor_coordinates):
+            center_coordinates = (
+                center_coordinates[0] + self.text_offset[0],
+                center_coordinates[1] + self.text_offset[1],
+            )
+
             wrapped_lines = wrap_text(label, self.max_line_length)
 
             # Calculate the total text height and maximum width
@@ -1409,7 +1550,7 @@ class RichLabelAnnotator(_BaseLabelAnnotator):
             height_padded = int(total_height + 2 * self.text_padding)
 
             text_background_xyxy = resolve_text_background_xyxy(
-                center_coordinates=tuple(center_coords),
+                center_coordinates=center_coordinates,
                 text_wh=(width_padded, height_padded),
                 position=self.text_anchor,
             )
@@ -1525,7 +1666,7 @@ class IconAnnotator(BaseAnnotator):
         self.position = icon_position
         self.offset_xy = offset_xy
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self, scene: ImageType, detections: Detections, icon_path: str | list[str]
     ) -> ImageType:
@@ -1614,7 +1755,7 @@ class BlurAnnotator(BaseAnnotator):
         """
         self.kernel_size: int = kernel_size
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -1681,6 +1822,7 @@ class TraceAnnotator(BaseAnnotator):
         position: Position = Position.CENTER,
         trace_length: int = 30,
         thickness: int = 2,
+        smooth: bool = False,
         color_lookup: ColorLookup = ColorLookup.CLASS,
     ):
         """
@@ -1692,15 +1834,17 @@ class TraceAnnotator(BaseAnnotator):
             trace_length (int): The maximum length of the trace in terms of historical
                 points. Defaults to `30`.
             thickness (int): The thickness of the trace lines. Defaults to `2`.
+            smooth (bool): Smooth the trace lines.
             color_lookup (ColorLookup): Strategy for mapping colors to annotations.
                 Options are `INDEX`, `CLASS`, `TRACK`.
         """
         self.color: Color | ColorPalette = color
         self.trace = Trace(max_size=trace_length, anchor=position)
         self.thickness = thickness
+        self.smooth = smooth
         self.color_lookup: ColorLookup = color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -1769,10 +1913,18 @@ class TraceAnnotator(BaseAnnotator):
                 else custom_color_lookup,
             )
             xy = self.trace.get(tracker_id=tracker_id)
+            spline_points = xy.astype(np.int32)
+
+            if len(xy) > 3 and self.smooth:
+                x, y = xy[:, 0], xy[:, 1]
+                tck, u = splprep([x, y], s=20)
+                x_new, y_new = splev(np.linspace(0, 1, 100), tck)
+                spline_points = np.stack([x_new, y_new], axis=1).astype(np.int32)
+
             if len(xy) > 1:
                 scene = cv2.polylines(
                     scene,
-                    [xy.astype(np.int32)],
+                    [spline_points],
                     False,
                     color=color.as_bgr(),
                     thickness=self.thickness,
@@ -1814,7 +1966,7 @@ class HeatMapAnnotator(BaseAnnotator):
         self.low_hue = low_hue
         self.heat_mask: npt.NDArray[np.float32] | None = None
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(self, scene: ImageType, detections: Detections) -> ImageType:
         """
         Annotates the scene with a heatmap based on the provided detections.
@@ -1896,7 +2048,7 @@ class PixelateAnnotator(BaseAnnotator):
         """
         self.pixel_size: int = pixel_size
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -1993,7 +2145,7 @@ class TriangleAnnotator(BaseAnnotator):
         self.outline_thickness: int = outline_thickness
         self.outline_color: Color | ColorPalette = outline_color
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -2105,7 +2257,7 @@ class RoundBoxAnnotator(BaseAnnotator):
             raise ValueError("roundness attribute must be float between (0, 1.0]")
         self.roundness: float = roundness
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -2245,7 +2397,7 @@ class PercentageBarAnnotator(BaseAnnotator):
             else int(0.15 * self.height)
         )
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -2426,7 +2578,7 @@ class CropAnnotator(BaseAnnotator):
         self.border_thickness: int = border_thickness
         self.border_color_lookup: ColorLookup = border_color_lookup
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self,
         scene: ImageType,
@@ -2575,7 +2727,7 @@ class BackgroundOverlayAnnotator(BaseAnnotator):
         self.opacity = opacity
         self.force_box = force_box
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(self, scene: ImageType, detections: Detections) -> ImageType:
         """
         Applies a colored overlay to the scene outside of the detected regions.
@@ -2673,7 +2825,7 @@ class ComparisonAnnotator:
         self.label_scale = label_scale
         self.text_thickness = int(self.label_scale + 1.2)
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(
         self, scene: ImageType, detections_1: Detections, detections_2: Detections
     ) -> ImageType:

supervision/annotators/utils.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import textwrap
 from enum import Enum
+from typing import Any
 
 import numpy as np
 
@@ -142,35 +143,45 @@ def resolve_color(
         detection_idx=detection_idx,
         color_lookup=color_lookup,
     )
-    if color_lookup == ColorLookup.TRACK and idx == PENDING_TRACK_ID:
+    if (
+        isinstance(color_lookup, ColorLookup)
+        and color_lookup == ColorLookup.TRACK
+        and idx == PENDING_TRACK_ID
+    ):
         return PENDING_TRACK_COLOR
     return get_color_by_index(color=color, idx=idx)
 
 
-def wrap_text(text: str, max_line_length=None) -> list[str]:
+def wrap_text(text: Any, max_line_length=None) -> list[str]:
     """
-    Wraps text to the specified maximum line length, respecting existing newlines.
-    Uses the textwrap library for robust text wrapping.
+    Wrap `text` to the specified maximum line length, respecting existing
+    newlines. Falls back to str() if `text` is not already a string.
 
     Args:
-        text (str): The text to wrap.
+        text (Any): The text (or object) to wrap.
+        max_line_length (int | None): Maximum width for each wrapped line.
 
     Returns:
-        List[str]: A list of text lines after wrapping.
+        list[str]: Wrapped lines.
     """
 
     if not text:
         return [""]
 
+    if not isinstance(text, str):
+        text = str(text)
+
     if max_line_length is None:
         return text.splitlines() or [""]
 
+    if max_line_length <= 0:
+        raise ValueError("max_line_length must be a positive integer")
+
     paragraphs = text.split("\n")
-    all_lines = []
+    all_lines: list[str] = []
 
     for paragraph in paragraphs:
-        if not paragraph:
-            # Keep empty lines
+        if paragraph == "":
             all_lines.append("")
             continue
 
@@ -182,12 +193,9 @@ def wrap_text(text: str, max_line_length=None) -> list[str]:
             drop_whitespace=True,
         )
 
-        if wrapped:
-            all_lines.extend(wrapped)
-        else:
-            all_lines.append("")
+        all_lines.extend(wrapped or [""])
 
-    return all_lines if all_lines else [""]
+    return all_lines or [""]
 
 
 def validate_labels(labels: list[str] | None, detections: Detections):

supervision/detection/core.py

@@ -40,6 +40,7 @@ from supervision.detection.utils.masks import calculate_masks_centroids
 from supervision.detection.vlm import (
     LMM,
     VLM,
+    from_deepseek_vl_2,
     from_florence_2,
     from_google_gemini_2_0,
     from_google_gemini_2_5,
@@ -830,6 +831,7 @@ class Detections:
         | Google Gemini 2.0   | `GOOGLE_GEMINI_2_0`  | detection               | `resolution_wh`             | `classes`           |
         | Google Gemini 2.5   | `GOOGLE_GEMINI_2_5`  | detection, segmentation | `resolution_wh`             | `classes`           |
         | Moondream           | `MOONDREAM`          | detection               | `resolution_wh`             |                     |
+        | DeepSeek-VL2        | `DEEPSEEK_VL_2`      | detection               | `resolution_wh`             | `classes`           |
 
         Args:
             lmm (Union[LMM, str]): The type of LMM (Large Multimodal Model) to use.
@@ -1117,6 +1119,47 @@ class Detections:
             # array([[1752.28,  818.82, 2165.72, 1229.14],
             #        [1908.01, 1346.67, 2585.99, 2024.11]])
             ```
+
+        !!! example "DeepSeek-VL2"
+
+
+            ??? tip "Prompt engineering"
+
+                To get the best results from DeepSeek-VL2, use optimized prompts that leverage
+                its object detection and visual grounding capabilities effectively.
+
+                **For general object detection, use the following user prompt:**
+
+                ```
+                <image>\\n<|ref|>The giraffe at the front<|/ref|>
+                ```
+
+                **For visual grounding, use the following user prompt:**
+
+                ```
+                <image>\\n<|grounding|>Detect the giraffes
+                ```
+
+            ```python
+            from PIL import Image
+            import supervision as sv
+
+            deepseek_vl2_result = "<|ref|>The giraffe at the back<|/ref|><|det|>[[580, 270, 999, 904]]<|/det|><|ref|>The giraffe at the front<|/ref|><|det|>[[26, 31, 632, 998]]<|/det|><|end▁of▁sentence|>"
+
+            detections = sv.Detections.from_vlm(
+                vlm=sv.VLM.DEEPSEEK_VL_2, result=deepseek_vl2_result, resolution_wh=image.size
+            )
+
+            detections.xyxy
+            # array([[ 420,  293,  724,  982],
+            #        [  18,   33,  458, 1084]])
+
+            detections.class_id
+            # array([0, 1])
+
+            detections.data
+            # {'class_name': array(['The giraffe at the back', 'The giraffe at the front'], dtype='<U24')}
+            ```
         """  # noqa: E501
 
         # filler logic mapping old from_lmm to new from_vlm
@@ -1124,6 +1167,7 @@ class Detections:
             LMM.PALIGEMMA: VLM.PALIGEMMA,
             LMM.FLORENCE_2: VLM.FLORENCE_2,
             LMM.QWEN_2_5_VL: VLM.QWEN_2_5_VL,
+            LMM.DEEPSEEK_VL_2: VLM.DEEPSEEK_VL_2,
             LMM.GOOGLE_GEMINI_2_0: VLM.GOOGLE_GEMINI_2_0,
             LMM.GOOGLE_GEMINI_2_5: VLM.GOOGLE_GEMINI_2_5,
         }
@@ -1164,6 +1208,7 @@ class Detections:
         | Google Gemini 2.0   | `GOOGLE_GEMINI_2_0`  | detection               | `resolution_wh`             | `classes`           |
         | Google Gemini 2.5   | `GOOGLE_GEMINI_2_5`  | detection, segmentation | `resolution_wh`             | `classes`           |
         | Moondream           | `MOONDREAM`          | detection               | `resolution_wh`             |                     |
+        | DeepSeek-VL2        | `DEEPSEEK_VL_2`      | detection               | `resolution_wh`             | `classes`           |
 
         Args:
             vlm (Union[VLM, str]): The type of VLM (Vision Language Model) to use.
@@ -1452,6 +1497,47 @@ class Detections:
             #        [1908.01, 1346.67, 2585.99, 2024.11]])
             ```
 
+        !!! example "DeepSeek-VL2"
+
+
+            ??? tip "Prompt engineering"
+
+                To get the best results from DeepSeek-VL2, use optimized prompts that leverage
+                its object detection and visual grounding capabilities effectively.
+
+                **For general object detection, use the following user prompt:**
+
+                ```
+                <image>\\n<|ref|>The giraffe at the front<|/ref|>
+                ```
+
+                **For visual grounding, use the following user prompt:**
+
+                ```
+                <image>\\n<|grounding|>Detect the giraffes
+                ```
+
+            ```python
+            from PIL import Image
+            import supervision as sv
+
+            deepseek_vl2_result = "<|ref|>The giraffe at the back<|/ref|><|det|>[[580, 270, 999, 904]]<|/det|><|ref|>The giraffe at the front<|/ref|><|det|>[[26, 31, 632, 998]]<|/det|><|end▁of▁sentence|>"
+
+            detections = sv.Detections.from_vlm(
+                vlm=sv.VLM.DEEPSEEK_VL_2, result=deepseek_vl2_result, resolution_wh=image.size
+            )
+
+            detections.xyxy
+            # array([[ 420,  293,  724,  982],
+            #        [  18,   33,  458, 1084]])
+
+            detections.class_id
+            # array([0, 1])
+
+            detections.data
+            # {'class_name': array(['The giraffe at the back', 'The giraffe at the front'], dtype='<U24')}
+            ```
+
         """  # noqa: E501
 
         vlm = validate_vlm_parameters(vlm, result, kwargs)
@@ -1466,6 +1552,11 @@ class Detections:
             data = {CLASS_NAME_DATA_FIELD: class_name}
             return cls(xyxy=xyxy, class_id=class_id, data=data)
 
+        if vlm == VLM.DEEPSEEK_VL_2:
+            xyxy, class_id, class_name = from_deepseek_vl_2(result, **kwargs)
+            data = {CLASS_NAME_DATA_FIELD: class_name}
+            return cls(xyxy=xyxy, class_id=class_id, data=data)
+
         if vlm == VLM.FLORENCE_2:
             xyxy, labels, mask, xyxyxyxy = from_florence_2(result, **kwargs)
             if len(xyxy) == 0:

supervision/detection/utils/__init__.py

@@ -0,0 +1,3 @@
+from supervision.detection.utils.iou_and_nms import box_iou_batch
+
+__all__ = ["box_iou_batch"]

supervision/detection/utils/converters.py

@@ -229,6 +229,70 @@ def mask_to_xyxy(masks: np.ndarray) -> np.ndarray:
     return xyxy
 
 
+def xyxy_to_mask(boxes: np.ndarray, resolution_wh: tuple[int, int]) -> np.ndarray:
+    """
+    Converts a 2D `np.ndarray` of bounding boxes into a 3D `np.ndarray` of bool masks.
+
+    Parameters:
+        boxes (np.ndarray): A 2D `np.ndarray` of shape `(N, 4)`
+            containing bounding boxes `(x_min, y_min, x_max, y_max)`
+        resolution_wh (Tuple[int, int]): A tuple `(width, height)` specifying
+            the resolution of the output masks
+
+    Returns:
+        np.ndarray: A 3D `np.ndarray` of shape `(N, height, width)`
+            containing 2D bool masks for each bounding box
+
+    Examples:
+        ```python
+        import numpy as np
+        import supervision as sv
+
+        boxes = np.array([[0, 0, 2, 2]])
+
+        sv.xyxy_to_mask(boxes, (5, 5))
+        # array([
+        #     [[ True,  True,  True, False, False],
+        #      [ True,  True,  True, False, False],
+        #      [ True,  True,  True, False, False],
+        #      [False, False, False, False, False],
+        #      [False, False, False, False, False]]
+        # ])
+
+        boxes = np.array([[0, 0, 1, 1], [3, 3, 4, 4]])
+
+        sv.xyxy_to_mask(boxes, (5, 5))
+        # array([
+        #     [[ True,  True, False, False, False],
+        #      [ True,  True, False, False, False],
+        #      [False, False, False, False, False],
+        #      [False, False, False, False, False],
+        #      [False, False, False, False, False]],
+        #
+        #     [[False, False, False, False, False],
+        #      [False, False, False, False, False],
+        #      [False, False, False, False, False],
+        #      [False, False, False,  True,  True],
+        #      [False, False, False,  True,  True]]
+        # ])
+        ```
+    """
+    width, height = resolution_wh
+    n = boxes.shape[0]
+    masks = np.zeros((n, height, width), dtype=bool)
+
+    for i, (x_min, y_min, x_max, y_max) in enumerate(boxes):
+        x_min = max(0, int(x_min))
+        y_min = max(0, int(y_min))
+        x_max = min(width - 1, int(x_max))
+        y_max = min(height - 1, int(y_max))
+
+        if x_max >= x_min and y_max >= y_min:
+            masks[i, y_min : y_max + 1, x_min : x_max + 1] = True
+
+    return masks
+
+
 def mask_to_polygons(mask: np.ndarray) -> list[np.ndarray]:
     """
     Converts a binary mask to a list of polygons.

supervision/detection/utils/internal.py

@@ -271,7 +271,6 @@ def merge_metadata(metadata_list: list[dict[str, Any]]) -> dict[str, Any]:
                     "{type(value)}, {type(other_value)}."
                 )
             else:
-                print("hm")
                 if merged_metadata[key] != value:
                     raise ValueError(f"Conflicting metadata for key: '{key}'.")
 

supervision/detection/utils/vlms.py

@@ -0,0 +1,105 @@
+from __future__ import annotations
+
+
+def edit_distance(string_1: str, string_2: str, case_sensitive: bool = True) -> int:
+    """
+    Calculates the minimum number of single-character edits required
+    to transform one string into another. Allowed operations are insertion,
+    deletion, and substitution.
+
+    Args:
+        string_1 (str): The source string to be transformed.
+        string_2 (str): The target string to transform into.
+        case_sensitive (bool, optional): Whether comparison should be case-sensitive.
+            Defaults to True.
+
+    Returns:
+        int: The minimum number of edits required to convert `string_1`
+        into `string_2`.
+
+    Examples:
+        ```python
+        import supervision as sv
+
+        sv.edit_distance("hello", "hello")
+        # 0
+
+        sv.edit_distance("Test", "test", case_sensitive=True)
+        # 1
+
+        sv.edit_distance("abc", "xyz")
+        # 3
+
+        sv.edit_distance("hello", "")
+        # 5
+
+        sv.edit_distance("", "")
+        # 0
+
+        sv.edit_distance("hello world", "helloworld")
+        # 1
+        ```
+    """
+    if not case_sensitive:
+        string_1 = string_1.lower()
+        string_2 = string_2.lower()
+
+    if len(string_1) < len(string_2):
+        string_1, string_2 = string_2, string_1
+
+    prev_row = list(range(len(string_2) + 1))
+    curr_row = [0] * (len(string_2) + 1)
+
+    for i in range(1, len(string_1) + 1):
+        curr_row[0] = i
+        for j in range(1, len(string_2) + 1):
+            if string_1[i - 1] == string_2[j - 1]:
+                substitution_cost = 0
+            else:
+                substitution_cost = 1
+            curr_row[j] = min(
+                prev_row[j] + 1,
+                curr_row[j - 1] + 1,
+                prev_row[j - 1] + substitution_cost,
+            )
+        prev_row, curr_row = curr_row, prev_row
+
+    return prev_row[len(string_2)]
+
+
+def fuzzy_match_index(
+    candidates: list[str],
+    query: str,
+    threshold: int,
+    case_sensitive: bool = True,
+) -> int | None:
+    """
+    Searches for the first string in `candidates` whose edit distance
+    to `query` is less than or equal to `threshold`.
+
+    Args:
+        candidates (list[str]): List of strings to search.
+        query (str): String to compare against the candidates.
+        threshold (int): Maximum allowed edit distance for a match.
+        case_sensitive (bool, optional): Whether matching should be case-sensitive.
+
+    Returns:
+        Optional[int]: Index of the first matching string in candidates,
+            or None if no match is found.
+
+    Examples:
+        ```python
+        fuzzy_match_index(["cat", "dog", "rat"], "dat", threshold=1)
+        # 0
+
+        fuzzy_match_index(["alpha", "beta", "gamma"], "bata", threshold=1)
+        # 1
+
+        fuzzy_match_index(["one", "two", "three"], "ten", threshold=2)
+        # None
+        ```
+    """
+    for idx, candidate in enumerate(candidates):
+        if edit_distance(candidate, query, case_sensitive=case_sensitive) <= threshold:
+            return idx
+    return None

supervision/detection/vlm.py

@@ -36,6 +36,7 @@ class LMM(Enum):
     PALIGEMMA = "paligemma"
     FLORENCE_2 = "florence_2"
     QWEN_2_5_VL = "qwen_2_5_vl"
+    DEEPSEEK_VL_2 = "deepseek_vl_2"
     GOOGLE_GEMINI_2_0 = "gemini_2_0"
     GOOGLE_GEMINI_2_5 = "gemini_2_5"
     MOONDREAM = "moondream"
@@ -76,6 +77,7 @@ class VLM(Enum):
     PALIGEMMA = "paligemma"
     FLORENCE_2 = "florence_2"
     QWEN_2_5_VL = "qwen_2_5_vl"
+    DEEPSEEK_VL_2 = "deepseek_vl_2"
     GOOGLE_GEMINI_2_0 = "gemini_2_0"
     GOOGLE_GEMINI_2_5 = "gemini_2_5"
     MOONDREAM = "moondream"
@@ -104,6 +106,7 @@ RESULT_TYPES: dict[VLM, type] = {
     VLM.PALIGEMMA: str,
     VLM.FLORENCE_2: dict,
     VLM.QWEN_2_5_VL: str,
+    VLM.DEEPSEEK_VL_2: str,
     VLM.GOOGLE_GEMINI_2_0: str,
     VLM.GOOGLE_GEMINI_2_5: str,
     VLM.MOONDREAM: dict,
@@ -113,6 +116,7 @@ REQUIRED_ARGUMENTS: dict[VLM, list[str]] = {
     VLM.PALIGEMMA: ["resolution_wh"],
     VLM.FLORENCE_2: ["resolution_wh"],
     VLM.QWEN_2_5_VL: ["input_wh", "resolution_wh"],
+    VLM.DEEPSEEK_VL_2: ["resolution_wh"],
     VLM.GOOGLE_GEMINI_2_0: ["resolution_wh"],
     VLM.GOOGLE_GEMINI_2_5: ["resolution_wh"],
     VLM.MOONDREAM: ["resolution_wh"],
@@ -122,6 +126,7 @@ ALLOWED_ARGUMENTS: dict[VLM, list[str]] = {
     VLM.PALIGEMMA: ["resolution_wh", "classes"],
     VLM.FLORENCE_2: ["resolution_wh"],
     VLM.QWEN_2_5_VL: ["input_wh", "resolution_wh", "classes"],
+    VLM.DEEPSEEK_VL_2: ["resolution_wh", "classes"],
     VLM.GOOGLE_GEMINI_2_0: ["resolution_wh", "classes"],
     VLM.GOOGLE_GEMINI_2_5: ["resolution_wh", "classes"],
     VLM.MOONDREAM: ["resolution_wh"],
@@ -310,6 +315,79 @@ def from_qwen_2_5_vl(
     return xyxy, class_id, class_name
 
 
+def from_deepseek_vl_2(
+    result: str, resolution_wh: tuple[int, int], classes: list[str] | None = None
+) -> tuple[np.ndarray, np.ndarray | None, np.ndarray]:
+    """
+    Parse bounding boxes from deepseek-vl2-formatted text, scale them to the specified
+    resolution, and optionally filter by classes.
+
+    The DeepSeek-VL2 output typically contains pairs of <|ref|> ... <|/ref|> labels
+    and <|det|> ... <|/det|> bounding box definitions. Each <|det|> section may
+    contain one or more bounding boxes in the form [[x1, y1, x2, y2], [x1, y1, x2, y2], ...]
+    (scaled to 0..999). For example:
+
+    ```
+    <|ref|>The giraffe at the back<|/ref|><|det|>[[580, 270, 999, 904]]<|/det|><|ref|>The giraffe at the front<|/ref|><|det|>[[26, 31, 632, 998]]<|/det|><|end▁of▁sentence|>
+    ```
+
+    Args:
+        result: String containing deepseek-vl2-formatted locations and labels.
+        resolution_wh: Tuple (width, height) to which we scale the box coordinates.
+        classes: Optional list of valid class names. If provided, boxes and labels not
+            in this list are filtered out.
+
+    Returns:
+        xyxy (np.ndarray): An array of shape `(n, 4)` containing
+            the bounding boxes coordinates in format `[x1, y1, x2, y2]`.
+        class_id (Optional[np.ndarray]): An array of shape `(n,)` containing
+            the class indices for each bounding box (or `None` if classes is not
+            provided).
+        class_name (np.ndarray): An array of shape `(n,)` containing
+            the class labels for each bounding box.
+    """  # noqa: E501
+
+    width, height = resolution_wh
+    label_segments = re.findall(r"<\|ref\|>(.*?)<\|/ref\|>", result, flags=re.S)
+    detection_segments = re.findall(r"<\|det\|>(.*?)<\|/det\|>", result, flags=re.S)
+
+    if len(label_segments) != len(detection_segments):
+        raise ValueError(
+            f"Number of ref tags ({len(label_segments)}) "
+            f"and det tags ({len(detection_segments)}) in the result must be equal."
+        )
+
+    xyxy, class_name_list = [], []
+    for label, detection_blob in zip(label_segments, detection_segments):
+        current_class_name = label.strip()
+        for box in re.findall(r"\[(.*?)\]", detection_blob):
+            x1, y1, x2, y2 = map(float, box.strip("[]").split(","))
+            xyxy.append(
+                [
+                    (x1 / 999 * width),
+                    (y1 / 999 * height),
+                    (x2 / 999 * width),
+                    (y2 / 999 * height),
+                ]
+            )
+            class_name_list.append(current_class_name)
+
+    xyxy = np.array(xyxy, dtype=np.float32)
+    class_name = np.array(class_name_list)
+
+    if classes is not None:
+        mask = np.array([name in classes for name in class_name], dtype=bool)
+        xyxy = xyxy[mask]
+        class_name = class_name[mask]
+        class_id = np.array([classes.index(name) for name in class_name])
+    else:
+        unique_classes = sorted(list(set(class_name)))
+        class_to_id = {name: i for i, name in enumerate(unique_classes)}
+        class_id = np.array([class_to_id[name] for name in class_name])
+
+    return xyxy, class_id, class_name
+
+
 def from_florence_2(
     result: dict, resolution_wh: tuple[int, int]
 ) -> tuple[np.ndarray, np.ndarray | None, np.ndarray | None, np.ndarray | None]:

supervision/draw/base.py

@@ -0,0 +1,13 @@
+from typing import TypeVar
+
+import numpy as np
+from PIL import Image
+
+ImageType = TypeVar("ImageType", np.ndarray, Image.Image)
+"""
+An image of type `np.ndarray` or `PIL.Image.Image`.
+
+Unlike a `Union`, ensures the type remains consistent. If a function
+takes an `ImageType` argument and returns an `ImageType`, when you
+pass an `np.ndarray`, you will get an `np.ndarray` back.
+"""

supervision/draw/utils.py

@@ -346,28 +346,50 @@ def draw_image(
 
 def calculate_optimal_text_scale(resolution_wh: tuple[int, int]) -> float:
     """
-    Calculate font scale based on the resolution of an image.
+    Calculate optimal font scale based on image resolution. Adjusts font scale
+    proportionally to the smallest dimension of the given image resolution for
+    consistent readability.
 
-    Parameters:
-        resolution_wh (Tuple[int, int]): A tuple representing the width and height
-            of the image.
+    Args:
+        resolution_wh (tuple[int, int]): (width, height) of the image in pixels
 
     Returns:
-         float: The calculated font scale factor.
+        float: recommended font scale factor
+
+    Examples:
+        ```python
+        import supervision as sv
+
+        sv.calculate_optimal_text_scale((1920, 1080))
+        # 1.08
+        sv.calculate_optimal_text_scale((640, 480))
+        # 0.48
+        ```
     """
     return min(resolution_wh) * 1e-3
 
 
 def calculate_optimal_line_thickness(resolution_wh: tuple[int, int]) -> int:
     """
-    Calculate line thickness based on the resolution of an image.
+    Calculate optimal line thickness based on image resolution. Adjusts the line
+    thickness for readability depending on the smallest dimension of the provided
+    image resolution.
 
-    Parameters:
-        resolution_wh (Tuple[int, int]): A tuple representing the width and height
-            of the image.
+    Args:
+        resolution_wh (tuple[int, int]): (width, height) of the image in pixels
 
     Returns:
-        int: The calculated line thickness in pixels.
+        int: recommended line thickness in pixels
+
+    Examples:
+        ```python
+        import supervision as sv
+
+        sv.calculate_optimal_line_thickness((1920, 1080))
+        # 4
+        sv.calculate_optimal_line_thickness((640, 480))
+        # 2
+        ```
     """
     if min(resolution_wh) < 1080:
         return 2

supervision/key_points/annotators.py

@@ -6,14 +6,14 @@ from logging import warn
 import cv2
 import numpy as np
 
-from supervision.annotators.base import ImageType
 from supervision.detection.utils.boxes import pad_boxes, spread_out_boxes
+from supervision.draw.base import ImageType
 from supervision.draw.color import Color
 from supervision.draw.utils import draw_rounded_rectangle
 from supervision.geometry.core import Rect
-from supervision.keypoint.core import KeyPoints
-from supervision.keypoint.skeletons import SKELETONS_BY_VERTEX_COUNT
-from supervision.utils.conversion import ensure_cv2_image_for_annotation
+from supervision.key_points.core import KeyPoints
+from supervision.key_points.skeletons import SKELETONS_BY_VERTEX_COUNT
+from supervision.utils.conversion import ensure_cv2_image_for_class_method
 
 
 class BaseKeyPointAnnotator(ABC):
@@ -43,7 +43,7 @@ class VertexAnnotator(BaseKeyPointAnnotator):
         self.color = color
         self.radius = radius
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(self, scene: ImageType, key_points: KeyPoints) -> ImageType:
         """
         Annotates the given scene with skeleton vertices based on the provided key
@@ -120,7 +120,7 @@ class EdgeAnnotator(BaseKeyPointAnnotator):
         self.thickness = thickness
         self.edges = edges
 
-    @ensure_cv2_image_for_annotation
+    @ensure_cv2_image_for_class_method
     def annotate(self, scene: ImageType, key_points: KeyPoints) -> ImageType:
         """
         Annotates the given scene by drawing lines between specified key points to form

supervision/key_points/core.py

@@ -10,7 +10,7 @@ import numpy.typing as npt
 from supervision.config import CLASS_NAME_DATA_FIELD
 from supervision.detection.core import Detections
 from supervision.detection.utils.internal import get_data_item, is_data_equal
-from supervision.validators import validate_keypoints_fields
+from supervision.validators import validate_key_points_fields
 
 
 @dataclass
@@ -23,7 +23,7 @@ class KeyPoints:
 
     === "Ultralytics"
 
-        Use [`sv.KeyPoints.from_ultralytics`](/latest/keypoint/core/#supervision.keypoint.core.KeyPoints.from_ultralytics)
+        Use [`sv.KeyPoints.from_ultralytics`](/latest/keypoint/core/#supervision.key_points.core.KeyPoints.from_ultralytics)
         method, which accepts [YOLOv8-pose](https://docs.ultralytics.com/models/yolov8/), [YOLO11-pose](https://docs.ultralytics.com/models/yolo11/)
         [pose](https://docs.ultralytics.com/tasks/pose/) result.
 
@@ -41,7 +41,7 @@ class KeyPoints:
 
     === "Inference"
 
-        Use [`sv.KeyPoints.from_inference`](/latest/keypoint/core/#supervision.keypoint.core.KeyPoints.from_inference)
+        Use [`sv.KeyPoints.from_inference`](/latest/keypoint/core/#supervision.key_points.core.KeyPoints.from_inference)
         method, which accepts [Inference](https://inference.roboflow.com/) pose result.
 
         ```python
@@ -58,7 +58,7 @@ class KeyPoints:
 
     === "MediaPipe"
 
-        Use [`sv.KeyPoints.from_mediapipe`](/latest/keypoint/core/#supervision.keypoint.core.KeyPoints.from_mediapipe)
+        Use [`sv.KeyPoints.from_mediapipe`](/latest/keypoint/core/#supervision.key_points.core.KeyPoints.from_mediapipe)
         method, which accepts [MediaPipe](https://github.com/google-ai-edge/mediapipe)
         pose result.
 
@@ -89,10 +89,61 @@ class KeyPoints:
             pose_landmarker_result, (image_width, image_height))
         ```
 
+    === "Transformers"
+
+        Use [`sv.KeyPoints.from_transformers`](/latest/keypoint/core/#supervision.key_points.core.KeyPoints.from_transformers)
+        method, which accepts [ViTPose](https://huggingface.co/docs/transformers/en/model_doc/vitpose) result.
+
+        ```python
+        from PIL import Image
+        import requests
+        import supervision as sv
+        import torch
+        from transformers import (
+            AutoProcessor,
+            RTDetrForObjectDetection,
+            VitPoseForPoseEstimation,
+        )
+
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        image = Image.open(<SOURCE_IMAGE_PATH>)
+
+        DETECTION_MODEL_ID = "PekingU/rtdetr_r50vd_coco_o365"
+
+        detection_processor = AutoProcessor.from_pretrained(DETECTION_MODEL_ID, use_fast=True)
+        detection_model = RTDetrForObjectDetection.from_pretrained(DETECTION_MODEL_ID, device_map=DEVICE)
+
+        inputs = detection_processor(images=frame, return_tensors="pt").to(DEVICE)
+
+        with torch.no_grad():
+            outputs = detection_model(**inputs)
+
+        target_size = torch.tensor([(frame.height, frame.width)])
+        results = detection_processor.post_process_object_detection(
+            outputs, target_sizes=target_size, threshold=0.3)
+
+        detections = sv.Detections.from_transformers(results[0])
+        boxes = sv.xyxy_to_xywh(detections[detections.class_id == 0].xyxy)
+
+        POSE_ESTIMATION_MODEL_ID = "usyd-community/vitpose-base-simple"
+
+        pose_estimation_processor = AutoProcessor.from_pretrained(POSE_ESTIMATION_MODEL_ID)
+        pose_estimation_model = VitPoseForPoseEstimation.from_pretrained(
+            POSE_ESTIMATION_MODEL_ID, device_map=DEVICE)
+
+        inputs = pose_estimation_processor(frame, boxes=[boxes], return_tensors="pt").to(DEVICE)
+
+        with torch.no_grad():
+            outputs = pose_estimation_model(**inputs)
+
+        results = pose_estimation_processor.post_process_pose_estimation(outputs, boxes=[boxes])
+        key_point = sv.KeyPoints.from_transformers(results[0])
+        ```
+
     Attributes:
         xy (np.ndarray): An array of shape `(n, m, 2)` containing
             `n` detected objects, each composed of `m` equally-sized
-            sets of keypoints, where each point is `[x, y]`.
+            sets of key points, where each point is `[x, y]`.
         class_id (Optional[np.ndarray]): An array of shape
             `(n,)` containing the class ids of the detected objects.
         confidence (Optional[np.ndarray]): An array of shape
@@ -109,7 +160,7 @@ class KeyPoints:
     data: dict[str, npt.NDArray[Any] | list] = field(default_factory=dict)
 
     def __post_init__(self):
-        validate_keypoints_fields(
+        validate_key_points_fields(
             xy=self.xy,
             confidence=self.confidence,
             class_id=self.class_id,
@@ -514,13 +565,13 @@ class KeyPoints:
             return cls.empty()
 
     @classmethod
-    def from_transformers(cls, transfomers_results: Any) -> KeyPoints:
+    def from_transformers(cls, transformers_results: Any) -> KeyPoints:
         """
         Create a `sv.KeyPoints` object from the
         [Transformers](https://github.com/huggingface/transformers) inference result.
 
         Args:
-            transfomers_results (Any): The output of a
+            transformers_results (Any): The output of a
                 Transformers model containing instances with prediction data.
 
         Returns:
@@ -545,9 +596,9 @@ class KeyPoints:
             DETECTION_MODEL_ID = "PekingU/rtdetr_r50vd_coco_o365"
 
             detection_processor = AutoProcessor.from_pretrained(DETECTION_MODEL_ID, use_fast=True)
-            detection_model = RTDetrForObjectDetection.from_pretrained(DETECTION_MODEL_ID, device_map=DEVICE)
+            detection_model = RTDetrForObjectDetection.from_pretrained(DETECTION_MODEL_ID, device_map=device)
 
-            inputs = detection_processor(images=frame, return_tensors="pt").to(DEVICE)
+            inputs = detection_processor(images=frame, return_tensors="pt").to(device)
 
             with torch.no_grad():
                 outputs = detection_model(**inputs)
@@ -563,9 +614,9 @@ class KeyPoints:
 
             pose_estimation_processor = AutoProcessor.from_pretrained(POSE_ESTIMATION_MODEL_ID)
             pose_estimation_model = VitPoseForPoseEstimation.from_pretrained(
-                POSE_ESTIMATION_MODEL_ID, device_map=DEVICE)
+                POSE_ESTIMATION_MODEL_ID, device_map=device)
 
-            inputs = pose_estimation_processor(frame, boxes=[boxes], return_tensors="pt").to(DEVICE)
+            inputs = pose_estimation_processor(frame, boxes=[boxes], return_tensors="pt").to(device)
 
             with torch.no_grad():
                 outputs = pose_estimation_model(**inputs)
@@ -576,8 +627,8 @@ class KeyPoints:
 
         """  # noqa: E501 // docs
 
-        if "keypoints" in transfomers_results[0]:
-            if transfomers_results[0]["keypoints"].cpu().numpy().size == 0:
+        if "keypoints" in transformers_results[0]:
+            if transformers_results[0]["keypoints"].cpu().numpy().size == 0:
                 return cls.empty()
 
             result_data = [
@@ -585,7 +636,7 @@ class KeyPoints:
                     result["keypoints"].cpu().numpy(),
                     result["scores"].cpu().numpy(),
                 )
-                for result in transfomers_results
+                for result in transformers_results
             ]
 
             xy, scores = zip(*result_data)
@@ -599,55 +650,68 @@ class KeyPoints:
             return cls.empty()
 
     def __getitem__(
-        self, index: int | slice | list[int] | np.ndarray | str
-    ) -> KeyPoints | list | np.ndarray | None:
-        """
-        Get a subset of the `sv.KeyPoints` object or access an item from its data field.
-
-        When provided with an integer, slice, list of integers, or a numpy array, this
-        method returns a new `sv.KeyPoints` object that represents a subset of the
-        original `sv.KeyPoints`. When provided with a string, it accesses the
-        corresponding item in the data dictionary.
-
-        Args:
-            index (Union[int, slice, List[int], np.ndarray, str]): The index, indices,
-                or key to access a subset of the `sv.KeyPoints` or an item from the
-                data.
-
-        Returns:
-            A subset of the `sv.KeyPoints` object or an item from the data field.
-
-        Examples:
-            ```python
-            import supervision as sv
-
-            key_points = sv.KeyPoints()
-
-            # access the first keypoint using an integer index
-            key_points[0]
-
-            # access the first 10 keypoints using index slice
-            key_points[0:10]
-
-            # access selected keypoints using a list of indices
-            key_points[[0, 2, 4]]
-
-            # access keypoints with selected class_id
-            key_points[key_points.class_id == 0]
-
-            # access keypoints with confidence greater than 0.5
-            key_points[key_points.confidence > 0.5]
-            ```
-        """
+        self, index: int | slice | list[int] | np.ndarray | tuple | str
+    ) -> KeyPoints | np.ndarray | list | None:
         if isinstance(index, str):
             return self.data.get(index)
-        if isinstance(index, int):
-            index = [index]
+
+        if not isinstance(index, tuple):
+            index = (index, slice(None))
+
+        i, j = index
+
+        if isinstance(i, int):
+            i = [i]
+
+        if isinstance(i, list) and all(isinstance(x, bool) for x in i):
+            i = np.array(i)
+        if isinstance(j, list) and all(isinstance(x, bool) for x in j):
+            j = np.array(j)
+
+        if isinstance(i, np.ndarray) and i.dtype == bool:
+            i = np.flatnonzero(i)
+        if isinstance(j, np.ndarray) and j.dtype == bool:
+            j = np.flatnonzero(j)
+
+        if (
+            isinstance(i, (list, np.ndarray))
+            and isinstance(j, (list, np.ndarray))
+            and not np.isscalar(i)
+            and not np.isscalar(j)
+        ):
+            i, j = np.ix_(i, j)
+
+        xy_selected = self.xy[i, j]
+
+        conf_selected = self.confidence[i, j] if self.confidence is not None else None
+
+        class_id_selected = self.class_id[i] if self.class_id is not None else None
+
+        data_selected = get_data_item(self.data, i)
+
+        if xy_selected.ndim == 1:
+            xy_selected = xy_selected.reshape(1, 1, 2)
+            if conf_selected is not None:
+                conf_selected = conf_selected.reshape(1, 1)
+        elif xy_selected.ndim == 2:
+            if np.isscalar(index[0]) or (
+                isinstance(index[0], np.ndarray) and index[0].ndim == 0
+            ):
+                xy_selected = xy_selected[np.newaxis, ...]
+                if conf_selected is not None:
+                    conf_selected = conf_selected[np.newaxis, ...]
+            elif np.isscalar(index[1]) or (
+                isinstance(index[1], np.ndarray) and index[1].ndim == 0
+            ):
+                xy_selected = xy_selected[:, np.newaxis, :]
+                if conf_selected is not None:
+                    conf_selected = conf_selected[:, np.newaxis]
+
         return KeyPoints(
-            xy=self.xy[index],
-            confidence=self.confidence[index] if self.confidence is not None else None,
-            class_id=self.class_id[index] if self.class_id is not None else None,
-            data=get_data_item(self.data, index),
+            xy=xy_selected,
+            confidence=conf_selected,
+            class_id=class_id_selected,
+            data=data_selected,
         )
 
     def __setitem__(self, key: str, value: np.ndarray | list):
@@ -668,12 +732,12 @@ class KeyPoints:
             model = YOLO('yolov8s.pt')
 
             result = model(image)[0]
-            keypoints = sv.KeyPoints.from_ultralytics(result)
+            key_points = sv.KeyPoints.from_ultralytics(result)
 
-            keypoints['class_name'] = [
+            key_points['class_name'] = [
                  model.model.names[class_id]
                  for class_id
-                 in keypoints.class_id
+                 in key_points.class_id
              ]
             ```
         """
@@ -688,7 +752,7 @@ class KeyPoints:
     @classmethod
     def empty(cls) -> KeyPoints:
         """
-        Create an empty Keypoints object with no keypoints.
+        Create an empty KeyPoints object with no key points.
 
         Returns:
             An empty `sv.KeyPoints` object.
@@ -706,9 +770,9 @@ class KeyPoints:
         """
         Returns `True` if the `KeyPoints` object is considered empty.
         """
-        empty_keypoints = KeyPoints.empty()
-        empty_keypoints.data = self.data
-        return self == empty_keypoints
+        empty_key_points = KeyPoints.empty()
+        empty_key_points.data = self.data
+        return self == empty_key_points
 
     def as_detections(
         self, selected_keypoint_indices: Iterable[int] | None = None
@@ -716,21 +780,21 @@ class KeyPoints:
         """
         Convert a KeyPoints object to a Detections object. This
         approximates the bounding box of the detected object by
-        taking the bounding box that fits all keypoints.
+        taking the bounding box that fits all key points.
 
         Arguments:
             selected_keypoint_indices (Optional[Iterable[int]]): The
-                indices of the keypoints to include in the bounding box
-                calculation. This helps focus on a subset of keypoints,
-                e.g. when some are occluded. Captures all keypoints by default.
+                indices of the key points to include in the bounding box
+                calculation. This helps focus on a subset of key points,
+                e.g. when some are occluded. Captures all key points by default.
 
         Returns:
             detections (Detections): The converted detections object.
 
         Examples:
             ```python
-            keypoints = sv.KeyPoints.from_inference(...)
-            detections = keypoints.as_detections()
+            key_points = sv.KeyPoints.from_inference(...)
+            detections = key_points.as_detections()
             ```
         """
         if self.is_empty():

supervision/metrics/mean_average_precision.py

@@ -102,12 +102,12 @@ class MeanAveragePrecisionResult:
             f"maxDets=100 ] = {self.map50:.3f}\n"
             f"Average Precision (AP) @[ IoU=0.75      | area=   all | "
             f"maxDets=100 ] = {self.map75:.3f}\n"
-            f"Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] "
-            f"= {self.small_objects.map50_95:.3f}\n"
-            f"Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] \
-                = {self.medium_objects.map50_95:.3f}\n"
-            f"Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] \
-                = {self.large_objects.map50_95:.3f}"
+            f"Average Precision (AP) @[ IoU=0.50:0.95 | area= small | "
+            f"maxDets=100 ] = {self.small_objects.map50_95:.3f}\n"
+            f"Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | "
+            f"maxDets=100 ] = {self.medium_objects.map50_95:.3f}\n"
+            f"Average Precision (AP) @[ IoU=0.50:0.95 | area= large | "
+            f"maxDets=100 ] = {self.large_objects.map50_95:.3f}"
         )
 
     def to_pandas(self) -> pd.DataFrame:

supervision/utils/conversion.py

@@ -4,10 +4,10 @@ import cv2
 import numpy as np
 from PIL import Image
 
-from supervision.annotators.base import ImageType
+from supervision.draw.base import ImageType
 
 
-def ensure_cv2_image_for_annotation(annotate_func):
+def ensure_cv2_image_for_class_method(annotate_func):
     """
     Decorates `BaseAnnotator.annotate` implementations, converts scene to
     an image type used internally by the annotators, converts back when annotation
@@ -32,7 +32,7 @@ def ensure_cv2_image_for_annotation(annotate_func):
     return wrapper
 
 
-def ensure_cv2_image_for_processing(image_processing_fun):
+def ensure_cv2_image_for_standalone_function(image_processing_fun):
     """
     Decorates image processing functions that accept np.ndarray, converting `image` to
     np.ndarray, converts back when processing is complete.
@@ -55,7 +55,7 @@ def ensure_cv2_image_for_processing(image_processing_fun):
     return wrapper
 
 
-def ensure_pil_image_for_annotation(annotate_func):
+def ensure_pil_image_for_class_method(annotate_func):
     """
     Decorates image processing functions that accept np.ndarray, converting `image` to
     PIL image, converts back when processing is complete.

supervision/utils/image.py

@@ -1,12 +1,7 @@
 from __future__ import annotations
 
-import itertools
-import math
 import os
 import shutil
-from collections.abc import Callable
-from functools import partial
-from typing import Literal
 
 import cv2
 import numpy as np
@@ -14,74 +9,60 @@ import numpy.typing as npt
 
 from supervision.annotators.base import ImageType
 from supervision.draw.color import Color, unify_to_bgr
-from supervision.draw.utils import calculate_optimal_text_scale, draw_text
-from supervision.geometry.core import Point
 from supervision.utils.conversion import (
-    cv2_to_pillow,
-    ensure_cv2_image_for_processing,
-    images_to_cv2,
+    ensure_cv2_image_for_standalone_function,
 )
-from supervision.utils.iterables import create_batches, fill
-
-RelativePosition = Literal["top", "bottom"]
-
-MAX_COLUMNS_FOR_SINGLE_ROW_GRID = 3
+from supervision.utils.internal import deprecated
 
 
-@ensure_cv2_image_for_processing
+@ensure_cv2_image_for_standalone_function
 def crop_image(
     image: ImageType,
     xyxy: npt.NDArray[int] | list[int] | tuple[int, int, int, int],
 ) -> ImageType:
     """
-    Crops the given image based on the given bounding box.
+    Crop image based on bounding box coordinates.
 
     Args:
-        image (ImageType): The image to be cropped. `ImageType` is a flexible type,
-            accepting either `numpy.ndarray` or `PIL.Image.Image`.
-        xyxy (Union[np.ndarray, List[int], Tuple[int, int, int, int]]): A bounding box
-            coordinates in the format `(x_min, y_min, x_max, y_max)`, accepted as either
-            a `numpy.ndarray`, a `list`, or a `tuple`.
+        image (`numpy.ndarray` or `PIL.Image.Image`): The image to crop.
+        xyxy (`numpy.array`, `list[int]`, or `tuple[int, int, int, int]`):
+            Bounding box coordinates in `(x_min, y_min, x_max, y_max)` format.
 
     Returns:
-        (ImageType): The cropped image. The type is determined by the input type and
-            may be either a `numpy.ndarray` or `PIL.Image.Image`.
-
-    === "OpenCV"
+        (`numpy.ndarray` or `PIL.Image.Image`): Cropped image matching input
+            type.
 
+    Examples:
         ```python
         import cv2
         import supervision as sv
 
-        image = cv2.imread(<SOURCE_IMAGE_PATH>)
+        image = cv2.imread("source.png")
         image.shape
         # (1080, 1920, 3)
 
-        xyxy = [200, 400, 600, 800]
+        xyxy = (400, 400, 800, 800)
         cropped_image = sv.crop_image(image=image, xyxy=xyxy)
         cropped_image.shape
         # (400, 400, 3)
         ```
 
-    === "Pillow"
-
         ```python
         from PIL import Image
         import supervision as sv
 
-        image = Image.open(<SOURCE_IMAGE_PATH>)
+        image = Image.open("source.png")
         image.size
         # (1920, 1080)
 
-        xyxy = [200, 400, 600, 800]
+        xyxy = (400, 400, 800, 800)
         cropped_image = sv.crop_image(image=image, xyxy=xyxy)
         cropped_image.size
         # (400, 400)
         ```
 
-    ![crop_image](https://media.roboflow.com/supervision-docs/crop-image.png){ align=center width="800" }
+    ![crop-image](https://media.roboflow.com/supervision-docs/supervision-docs-crop-image-2.png){ align=center width="1000" }
     """  # noqa E501 // docs
-
     if isinstance(xyxy, (list, tuple)):
         xyxy = np.array(xyxy)
     xyxy = np.round(xyxy).astype(int)
@@ -89,31 +70,28 @@ def crop_image(
     return image[y_min:y_max, x_min:x_max]
 
 
-@ensure_cv2_image_for_processing
+@ensure_cv2_image_for_standalone_function
 def scale_image(image: ImageType, scale_factor: float) -> ImageType:
     """
-    Scales the given image based on the given scale factor.
+    Scale image by given factor. Scale factor > 1.0 zooms in, < 1.0 zooms out.
 
     Args:
-        image (ImageType): The image to be scaled. `ImageType` is a flexible type,
-            accepting either `numpy.ndarray` or `PIL.Image.Image`.
-        scale_factor (float): The factor by which the image will be scaled. Scale
-            factor > `1.0` zooms in, < `1.0` zooms out.
+        image (`numpy.ndarray` or `PIL.Image.Image`): The image to scale.
+        scale_factor (`float`): Factor by which to scale the image.
 
     Returns:
-        (ImageType): The scaled image. The type is determined by the input type and
-            may be either a `numpy.ndarray` or `PIL.Image.Image`.
+        (`numpy.ndarray` or `PIL.Image.Image`): Scaled image matching input
+            type.
 
     Raises:
-        ValueError: If the scale factor is non-positive.
-
-    === "OpenCV"
+        ValueError: If scale factor is non-positive.
 
+    Examples:
         ```python
         import cv2
         import supervision as sv
 
-        image = cv2.imread(<SOURCE_IMAGE_PATH>)
+        image = cv2.imread("source.png")
         image.shape
         # (1080, 1920, 3)
 
@@ -122,13 +100,11 @@ def scale_image(image: ImageType, scale_factor: float) -> ImageType:
         # (540, 960, 3)
         ```
 
-    === "Pillow"
-
         ```python
         from PIL import Image
         import supervision as sv
 
-        image = Image.open(<SOURCE_IMAGE_PATH>)
+        image = Image.open("source.png")
         image.size
         # (1920, 1080)
 
@@ -136,7 +112,9 @@ def scale_image(image: ImageType, scale_factor: float) -> ImageType:
         scaled_image.size
         # (960, 540)
         ```
-    """
+
+    ![scale-image](https://media.roboflow.com/supervision-docs/supervision-docs-scale-image-2.png){ align=center width="1000" }
+    """  # noqa E501 // docs
     if scale_factor <= 0:
         raise ValueError("Scale factor must be positive.")
 
@@ -146,35 +124,31 @@ def scale_image(image: ImageType, scale_factor: float) -> ImageType:
     return cv2.resize(image, (width_new, height_new), interpolation=cv2.INTER_LINEAR)
 
 
-@ensure_cv2_image_for_processing
+@ensure_cv2_image_for_standalone_function
 def resize_image(
     image: ImageType,
     resolution_wh: tuple[int, int],
     keep_aspect_ratio: bool = False,
 ) -> ImageType:
     """
-    Resizes the given image to a specified resolution. Can maintain the original aspect
-    ratio or resize directly to the desired dimensions.
+    Resize image to specified resolution. Can optionally maintain aspect ratio.
 
     Args:
-        image (ImageType): The image to be resized. `ImageType` is a flexible type,
-            accepting either `numpy.ndarray` or `PIL.Image.Image`.
-        resolution_wh (Tuple[int, int]): The target resolution as
-            `(width, height)`.
-        keep_aspect_ratio (bool): Flag to maintain the image's original
-            aspect ratio. Defaults to `False`.
+        image (`numpy.ndarray` or `PIL.Image.Image`): The image to resize.
+        resolution_wh (`tuple[int, int]`): Target resolution as `(width, height)`.
+        keep_aspect_ratio (`bool`): Flag to maintain original aspect ratio.
+            Defaults to `False`.
 
     Returns:
-        (ImageType): The resized image. The type is determined by the input type and
-            may be either a `numpy.ndarray` or `PIL.Image.Image`.
-
-    === "OpenCV"
+        (`numpy.ndarray` or `PIL.Image.Image`): Resized image matching input
+            type.
 
+    Examples:
         ```python
         import cv2
         import supervision as sv
 
-        image = cv2.imread(<SOURCE_IMAGE_PATH>)
+        image = cv2.imread("source.png")
         image.shape
         # (1080, 1920, 3)
 
@@ -185,13 +159,11 @@ def resize_image(
         # (562, 1000, 3)
         ```
 
-    === "Pillow"
-
         ```python
         from PIL import Image
         import supervision as sv
 
-        image = Image.open(<SOURCE_IMAGE_PATH>)
+        image = Image.open("source.png")
         image.size
         # (1920, 1080)
 
@@ -202,7 +174,7 @@ def resize_image(
         # (1000, 562)
         ```
 
-    ![resize_image](https://media.roboflow.com/supervision-docs/resize-image.png){ align=center width="800" }
+    ![resize-image](https://media.roboflow.com/supervision-docs/supervision-docs-resize-image-2.png){ align=center width="1000" }
     """  # noqa E501 // docs
     if keep_aspect_ratio:
         image_ratio = image.shape[1] / image.shape[0]
@@ -219,59 +191,58 @@ def resize_image(
     return cv2.resize(image, (width_new, height_new), interpolation=cv2.INTER_LINEAR)
 
 
-@ensure_cv2_image_for_processing
+@ensure_cv2_image_for_standalone_function
 def letterbox_image(
     image: ImageType,
     resolution_wh: tuple[int, int],
     color: tuple[int, int, int] | Color = Color.BLACK,
 ) -> ImageType:
     """
-    Resizes and pads an image to a specified resolution with a given color, maintaining
-    the original aspect ratio.
+    Resize image and pad with color to achieve desired resolution while
+    maintaining aspect ratio.
 
     Args:
-        image (ImageType): The image to be resized. `ImageType` is a flexible type,
-            accepting either `numpy.ndarray` or `PIL.Image.Image`.
-        resolution_wh (Tuple[int, int]): The target resolution as
-            `(width, height)`.
-        color (Union[Tuple[int, int, int], Color]): The color to pad with. If tuple
-            provided it should be in BGR format.
+        image (`numpy.ndarray` or `PIL.Image.Image`): The image to resize and pad.
+        resolution_wh (`tuple[int, int]`): Target resolution as `(width, height)`.
+        color (`tuple[int, int, int]` or `Color`): Padding color. If tuple, should
+            be in BGR format. Defaults to `Color.BLACK`.
 
     Returns:
-        (ImageType): The resized image. The type is determined by the input type and
-            may be either a `numpy.ndarray` or `PIL.Image.Image`.
-
-    === "OpenCV"
+        (`numpy.ndarray` or `PIL.Image.Image`): Letterboxed image matching input
+            type.
 
+    Examples:
         ```python
         import cv2
         import supervision as sv
 
-        image = cv2.imread(<SOURCE_IMAGE_PATH>)
+        image = cv2.imread("source.png")
         image.shape
         # (1080, 1920, 3)
 
-        letterboxed_image = sv.letterbox_image(image=image, resolution_wh=(1000, 1000))
+        letterboxed_image = sv.letterbox_image(
+            image=image, resolution_wh=(1000, 1000)
+        )
         letterboxed_image.shape
         # (1000, 1000, 3)
         ```
 
-    === "Pillow"
-
         ```python
         from PIL import Image
         import supervision as sv
 
-        image = Image.open(<SOURCE_IMAGE_PATH>)
+        image = Image.open("source.png")
         image.size
         # (1920, 1080)
 
-        letterboxed_image = sv.letterbox_image(image=image, resolution_wh=(1000, 1000))
+        letterboxed_image = sv.letterbox_image(
+            image=image, resolution_wh=(1000, 1000)
+        )
         letterboxed_image.size
         # (1000, 1000)
         ```
 
-    ![letterbox_image](https://media.roboflow.com/supervision-docs/letterbox-image.png){ align=center width="800" }
+    ![letterbox-image](https://media.roboflow.com/supervision-docs/supervision-docs-letterbox-image-2.png){ align=center width="1000" }
     """  # noqa E501 // docs
     assert isinstance(image, np.ndarray)
     color = unify_to_bgr(color=color)
@@ -302,37 +273,59 @@ def letterbox_image(
     return image_with_borders
 
 
+@deprecated(
+    "`overlay_image` function is deprecated and will be removed in "
+    "`supervision-0.32.0`. Use `draw_image` instead."
+)
 def overlay_image(
     image: npt.NDArray[np.uint8],
     overlay: npt.NDArray[np.uint8],
     anchor: tuple[int, int],
 ) -> npt.NDArray[np.uint8]:
     """
-    Places an image onto a scene at a given anchor point, handling cases where
-    the image's position is partially or completely outside the scene's bounds.
+    Overlay image onto scene at specified anchor point. Handles cases where
+    overlay position is partially or completely outside scene bounds.
 
     Args:
-        image (np.ndarray): The background scene onto which the image is placed.
-        overlay (np.ndarray): The image to be placed onto the scene.
-        anchor (Tuple[int, int]): The `(x, y)` coordinates in the scene where the
-            top-left corner of the image will be placed.
+        image (`numpy.array`): Background scene with shape `(height, width, 3)`.
+        overlay (`numpy.array`): Image to overlay with shape
+            `(height, width, 3)` or `(height, width, 4)`.
+        anchor (`tuple[int, int]`): Coordinates `(x, y)` where top-left corner
+            of overlay will be placed.
 
     Returns:
-        (np.ndarray): The result image with overlay.
+        (`numpy.array`): Scene with overlay applied, shape `(height, width, 3)`.
 
     Examples:
-        ```python
+        ```
         import cv2
         import numpy as np
         import supervision as sv
 
-        image = cv2.imread(<SOURCE_IMAGE_PATH>)
+        image = cv2.imread("source.png")
         overlay = np.zeros((400, 400, 3), dtype=np.uint8)
-        result_image = sv.overlay_image(image=image, overlay=overlay, anchor=(200, 400))
+        overlay[:] = (0, 255, 0)  # Green overlay
+
+        result_image = sv.overlay_image(
+            image=image, overlay=overlay, anchor=(200, 400)
+        )
+        cv2.imwrite("target.png", result_image)
         ```
 
-    ![overlay_image](https://media.roboflow.com/supervision-docs/overlay-image.png){ align=center width="800" }
-    """  # noqa E501 // docs
+        ```
+        import cv2
+        import numpy as np
+        import supervision as sv
+
+        image = cv2.imread("source.png")
+        overlay = cv2.imread("overlay.png", cv2.IMREAD_UNCHANGED)
+
+        result_image = sv.overlay_image(
+            image=image, overlay=overlay, anchor=(100, 100)
+        )
+        cv2.imwrite("target.png", result_image)
+        ```
+    """
     scene_height, scene_width = image.shape[:2]
     image_height, image_width = overlay.shape[:2]
     anchor_x, anchor_y = anchor
@@ -371,6 +364,102 @@ def overlay_image(
     return image
 
 
+@ensure_cv2_image_for_standalone_function
+def tint_image(
+    image: ImageType,
+    color: Color = Color.BLACK,
+    opacity: float = 0.5,
+) -> ImageType:
+    """
+    Tint image with solid color overlay at specified opacity.
+
+    Args:
+        image (`numpy.ndarray` or `PIL.Image.Image`): The image to tint.
+        color (`Color`): Overlay tint color. Defaults to `Color.BLACK`.
+        opacity (`float`): Blend ratio between overlay and image (0.0-1.0).
+            Defaults to `0.5`.
+
+    Returns:
+        (`numpy.ndarray` or `PIL.Image.Image`): Tinted image matching input
+            type.
+
+    Raises:
+        ValueError: If opacity is outside range [0.0, 1.0].
+
+    Examples:
+        ```python
+        import cv2
+        import supervision as sv
+
+        image = cv2.imread("source.png")
+        tinted_image = sv.tint_image(
+            image=image, color=sv.Color.ROBOFLOW, opacity=0.5
+        )
+        cv2.imwrite("target.png", tinted_image)
+        ```
+
+        ```python
+        from PIL import Image
+        import supervision as sv
+
+        image = Image.open("source.png")
+        tinted_image = sv.tint_image(
+            image=image, color=sv.Color.ROBOFLOW, opacity=0.5
+        )
+        tinted_image.save("target.png")
+        ```
+
+    ![tint-image](https://media.roboflow.com/supervision-docs/supervision-docs-tint-image-2.png){ align=center width="1000" }
+    """  # noqa E501 // docs
+    if not 0.0 <= opacity <= 1.0:
+        raise ValueError("opacity must be between 0.0 and 1.0")
+
+    overlay = np.full_like(image, fill_value=color.as_bgr(), dtype=image.dtype)
+    cv2.addWeighted(
+        src1=overlay, alpha=opacity, src2=image, beta=1 - opacity, gamma=0, dst=image
+    )
+    return image
+
+
+@ensure_cv2_image_for_standalone_function
+def grayscale_image(image: ImageType) -> ImageType:
+    """
+    Convert image to 3-channel grayscale. Luminance channel is broadcast to
+    all three channels for compatibility with color-based drawing helpers.
+
+    Args:
+        image (`numpy.ndarray` or `PIL.Image.Image`): The image to convert to
+            grayscale.
+
+    Returns:
+        (`numpy.ndarray` or `PIL.Image.Image`): 3-channel grayscale image
+            matching input type.
+
+    Examples:
+        ```python
+        import cv2
+        import supervision as sv
+
+        image = cv2.imread("source.png")
+        grayscale_image = sv.grayscale_image(image=image)
+        cv2.imwrite("target.png", grayscale_image)
+        ```
+
+        ```python
+        from PIL import Image
+        import supervision as sv
+
+        image = Image.open("source.png")
+        grayscale_image = sv.grayscale_image(image=image)
+        grayscale_image.save("target.png")
+        ```
+
+    ![grayscale-image](https://media.roboflow.com/supervision-docs/supervision-docs-grayscale-image-2.png){ align=center width="1000" }
+    """  # noqa E501 // docs
+    grayscaled = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+    return cv2.cvtColor(grayscaled, cv2.COLOR_GRAY2BGR)
+
+
 class ImageSink:
     def __init__(
         self,
@@ -379,27 +468,64 @@ class ImageSink:
         image_name_pattern: str = "image_{:05d}.png",
     ):
         """
-        Initialize a context manager for saving images.
+        Initialize context manager for saving images to directory.
 
         Args:
-            target_dir_path (str): The target directory where images will be saved.
-            overwrite (bool): Whether to overwrite the existing directory.
-                Defaults to False.
-            image_name_pattern (str): The image file name pattern.
-                Defaults to "image_{:05d}.png".
+            target_dir_path (`str`): Target directory path where images will be
+                saved.
+            overwrite (`bool`): Whether to overwrite existing directory.
+                Defaults to `False`.
+            image_name_pattern (`str`): File name pattern for saved images.
+                Defaults to `"image_{:05d}.png"`.
 
         Examples:
             ```python
             import supervision as sv
 
-            frames_generator = sv.get_video_frames_generator(<SOURCE_VIDEO_PATH>, stride=2)
+            frames_generator = sv.get_video_frames_generator(
+                "source.mp4", stride=2
+            )
 
-            with sv.ImageSink(target_dir_path=<TARGET_CROPS_DIRECTORY>) as sink:
+            with sv.ImageSink(target_dir_path="output_frames") as sink:
                 for image in frames_generator:
                     sink.save_image(image=image)
-            ```
-        """  # noqa E501 // docs
 
+            # Directory structure:
+            # output_frames/
+            # ├── image_00000.png
+            # ├── image_00001.png
+            # ├── image_00002.png
+            # └── image_00003.png
+            ```
+
+            ```python
+            import cv2
+            import supervision as sv
+
+            image = cv2.imread("source.png")
+            crop_boxes = [
+                (  0,   0, 400, 400),
+                (400,   0, 800, 400),
+                (  0, 400, 400, 800),
+                (400, 400, 800, 800)
+            ]
+
+            with sv.ImageSink(
+                target_dir_path="image_crops",
+                overwrite=True
+            ) as sink:
+                for i, xyxy in enumerate(crop_boxes):
+                    crop = sv.crop_image(image=image, xyxy=xyxy)
+                    sink.save_image(image=crop, image_name=f"crop_{i}.png")
+
+            # Directory structure:
+            # image_crops/
+            # ├── crop_0.png
+            # ├── crop_1.png
+            # ├── crop_2.png
+            # └── crop_3.png
+            ```
+        """
         self.target_dir_path = target_dir_path
         self.overwrite = overwrite
         self.image_name_pattern = image_name_pattern
@@ -417,14 +543,14 @@ class ImageSink:
 
     def save_image(self, image: np.ndarray, image_name: str | None = None):
         """
-        Save a given image in the target directory.
+        Save image to target directory with optional custom filename.
 
         Args:
-            image (np.ndarray): The image to be saved. The image must be in BGR color
-                format.
-            image_name (Optional[str]): The name to use for the saved image.
-                If not provided, a name will be
-                generated using the `image_name_pattern`.
+            image (`numpy.array`): Image to save with shape `(height, width, 3)`
+                in BGR format.
+            image_name (`str` or `None`): Custom filename for saved image. If
+                `None`, generates name using `image_name_pattern`. Defaults to
+                `None`.
         """
         if image_name is None:
             image_name = self.image_name_pattern.format(self.image_count)
@@ -435,355 +561,3 @@ class ImageSink:
 
     def __exit__(self, exc_type, exc_value, exc_traceback):
         pass
-
-
-def create_tiles(
-    images: list[ImageType],
-    grid_size: tuple[int | None, int | None] | None = None,
-    single_tile_size: tuple[int, int] | None = None,
-    tile_scaling: Literal["min", "max", "avg"] = "avg",
-    tile_padding_color: tuple[int, int, int] | Color = Color.from_hex("#D9D9D9"),
-    tile_margin: int = 10,
-    tile_margin_color: tuple[int, int, int] | Color = Color.from_hex("#BFBEBD"),
-    return_type: Literal["auto", "cv2", "pillow"] = "auto",
-    titles: list[str | None] | None = None,
-    titles_anchors: Point | list[Point | None] | None = None,
-    titles_color: tuple[int, int, int] | Color = Color.from_hex("#262523"),
-    titles_scale: float | None = None,
-    titles_thickness: int = 1,
-    titles_padding: int = 10,
-    titles_text_font: int = cv2.FONT_HERSHEY_SIMPLEX,
-    titles_background_color: tuple[int, int, int] | Color = Color.from_hex("#D9D9D9"),
-    default_title_placement: RelativePosition = "top",
-) -> ImageType:
-    """
-    Creates tiles mosaic from input images, automating grid placement and
-    converting images to common resolution maintaining aspect ratio. It is
-    also possible to render text titles on tiles, using optional set of
-    parameters specifying text drawing (see parameters description).
-
-    Automated grid placement will try to maintain square shape of grid
-    (with size being the nearest integer square root of #images), up to two exceptions:
-    * if there are up to 3 images - images will be displayed in single row
-    * if square-grid placement causes last row to be empty - number of rows is trimmed
-        until last row has at least one image
-
-    Args:
-        images (List[ImageType]): Images to create tiles. Elements can be either
-            np.ndarray or PIL.Image, common representation will be agreed by the
-            function.
-        grid_size (Optional[Tuple[Optional[int], Optional[int]]]): Expected grid
-            size in format (n_rows, n_cols). If not given - automated grid placement
-            will be applied. One may also provide only one out of two elements of the
-            tuple - then grid will be created with either n_rows or n_cols fixed,
-            leaving the other dimension to be adjusted by the number of images
-        single_tile_size (Optional[Tuple[int, int]]): sizeof a single tile element
-            provided in (width, height) format. If not given - size of tile will be
-            automatically calculated based on `tile_scaling` parameter.
-        tile_scaling (Literal["min", "max", "avg"]): If `single_tile_size` is not
-            given - parameter will be used to calculate tile size - using
-            min / max / avg size of image provided in `images` list.
-        tile_padding_color (Union[Tuple[int, int, int], sv.Color]): Color to be used in
-            images letterbox procedure (while standardising tiles sizes) as a padding.
-            If tuple provided - should be BGR.
-        tile_margin (int): size of margin between tiles (in pixels)
-        tile_margin_color (Union[Tuple[int, int, int], sv.Color]): Color of tile margin.
-            If tuple provided - should be BGR.
-        return_type (Literal["auto", "cv2", "pillow"]): Parameter dictates the format of
-            return image. One may choose specific type ("cv2" or "pillow") to enforce
-            conversion. "auto" mode takes a majority vote between types of elements in
-            `images` list - resolving draws in favour of OpenCV format. "auto" can be
-            safely used when all input images are of the same type.
-        titles (Optional[List[Optional[str]]]): Optional titles to be added to tiles.
-            Elements of that list may be empty - then specific tile (in order presented
-            in `images` parameter) will not be filled with title. It is possible to
-            provide list of titles shorter than `images` - then remaining titles will
-            be assumed empty.
-        titles_anchors (Optional[Union[Point, List[Optional[Point]]]]): Parameter to
-            specify anchor points for titles. It is possible to specify anchor either
-            globally or for specific tiles (following order of `images`).
-            If not given (either globally, or for specific element of the list),
-            it will be calculated automatically based on `default_title_placement`.
-        titles_color (Union[Tuple[int, int, int], Color]): Color of titles text.
-            If tuple provided - should be BGR.
-        titles_scale (Optional[float]): Scale of titles. If not provided - value will
-            be calculated using `calculate_optimal_text_scale(...)`.
-        titles_thickness (int): Thickness of titles text.
-        titles_padding (int): Size of titles padding.
-        titles_text_font (int): Font to be used to render titles. Must be integer
-            constant representing OpenCV font.
-            (See docs: https://docs.opencv.org/4.x/d6/d6e/group__imgproc__draw.html)
-        titles_background_color (Union[Tuple[int, int, int], Color]): Color of title
-            text padding.
-        default_title_placement (Literal["top", "bottom"]): Parameter specifies title
-            anchor placement in case if explicit anchor is not provided.
-
-    Returns:
-        ImageType: Image with all input images located in tails grid. The output type is
-            determined by `return_type` parameter.
-
-    Raises:
-        ValueError: In case when input images list is empty, provided `grid_size` is too
-            small to fit all images, `tile_scaling` mode is invalid.
-    """
-    if len(images) == 0:
-        raise ValueError("Could not create image tiles from empty list of images.")
-    if return_type == "auto":
-        return_type = _negotiate_tiles_format(images=images)
-    tile_padding_color = unify_to_bgr(color=tile_padding_color)
-    tile_margin_color = unify_to_bgr(color=tile_margin_color)
-    images = images_to_cv2(images=images)
-    if single_tile_size is None:
-        single_tile_size = _aggregate_images_shape(images=images, mode=tile_scaling)
-    resized_images = [
-        letterbox_image(
-            image=i, resolution_wh=single_tile_size, color=tile_padding_color
-        )
-        for i in images
-    ]
-    grid_size = _establish_grid_size(images=images, grid_size=grid_size)
-    if len(images) > grid_size[0] * grid_size[1]:
-        raise ValueError(
-            f"Could not place {len(images)} in grid with size: {grid_size}."
-        )
-    if titles is not None:
-        titles = fill(sequence=titles, desired_size=len(images), content=None)
-    titles_anchors = (
-        [titles_anchors]
-        if not issubclass(type(titles_anchors), list)
-        else titles_anchors
-    )
-    titles_anchors = fill(
-        sequence=titles_anchors, desired_size=len(images), content=None
-    )
-    titles_color = unify_to_bgr(color=titles_color)
-    titles_background_color = unify_to_bgr(color=titles_background_color)
-    tiles = _generate_tiles(
-        images=resized_images,
-        grid_size=grid_size,
-        single_tile_size=single_tile_size,
-        tile_padding_color=tile_padding_color,
-        tile_margin=tile_margin,
-        tile_margin_color=tile_margin_color,
-        titles=titles,
-        titles_anchors=titles_anchors,
-        titles_color=titles_color,
-        titles_scale=titles_scale,
-        titles_thickness=titles_thickness,
-        titles_padding=titles_padding,
-        titles_text_font=titles_text_font,
-        titles_background_color=titles_background_color,
-        default_title_placement=default_title_placement,
-    )
-    if return_type == "pillow":
-        tiles = cv2_to_pillow(image=tiles)
-    return tiles
-
-
-def _negotiate_tiles_format(images: list[ImageType]) -> Literal["cv2", "pillow"]:
-    number_of_np_arrays = sum(issubclass(type(i), np.ndarray) for i in images)
-    if number_of_np_arrays >= (len(images) // 2):
-        return "cv2"
-    return "pillow"
-
-
-def _calculate_aggregated_images_shape(
-    images: list[np.ndarray], aggregator: Callable[[list[int]], float]
-) -> tuple[int, int]:
-    height = round(aggregator([i.shape[0] for i in images]))
-    width = round(aggregator([i.shape[1] for i in images]))
-    return width, height
-
-
-SHAPE_AGGREGATION_FUN = {
-    "min": partial(_calculate_aggregated_images_shape, aggregator=np.min),
-    "max": partial(_calculate_aggregated_images_shape, aggregator=np.max),
-    "avg": partial(_calculate_aggregated_images_shape, aggregator=np.average),
-}
-
-
-def _aggregate_images_shape(
-    images: list[np.ndarray], mode: Literal["min", "max", "avg"]
-) -> tuple[int, int]:
-    if mode not in SHAPE_AGGREGATION_FUN:
-        raise ValueError(
-            f"Could not aggregate images shape - provided unknown mode: {mode}. "
-            f"Supported modes: {list(SHAPE_AGGREGATION_FUN.keys())}."
-        )
-    return SHAPE_AGGREGATION_FUN[mode](images)
-
-
-def _establish_grid_size(
-    images: list[np.ndarray], grid_size: tuple[int | None, int | None] | None
-) -> tuple[int, int]:
-    if grid_size is None or all(e is None for e in grid_size):
-        return _negotiate_grid_size(images=images)
-    if grid_size[0] is None:
-        return math.ceil(len(images) / grid_size[1]), grid_size[1]
-    if grid_size[1] is None:
-        return grid_size[0], math.ceil(len(images) / grid_size[0])
-    return grid_size
-
-
-def _negotiate_grid_size(images: list[np.ndarray]) -> tuple[int, int]:
-    if len(images) <= MAX_COLUMNS_FOR_SINGLE_ROW_GRID:
-        return 1, len(images)
-    nearest_sqrt = math.ceil(np.sqrt(len(images)))
-    proposed_columns = nearest_sqrt
-    proposed_rows = nearest_sqrt
-    while proposed_columns * (proposed_rows - 1) >= len(images):
-        proposed_rows -= 1
-    return proposed_rows, proposed_columns
-
-
-def _generate_tiles(
-    images: list[np.ndarray],
-    grid_size: tuple[int, int],
-    single_tile_size: tuple[int, int],
-    tile_padding_color: tuple[int, int, int],
-    tile_margin: int,
-    tile_margin_color: tuple[int, int, int],
-    titles: list[str | None] | None,
-    titles_anchors: list[Point | None],
-    titles_color: tuple[int, int, int],
-    titles_scale: float | None,
-    titles_thickness: int,
-    titles_padding: int,
-    titles_text_font: int,
-    titles_background_color: tuple[int, int, int],
-    default_title_placement: RelativePosition,
-) -> np.ndarray:
-    images = _draw_texts(
-        images=images,
-        titles=titles,
-        titles_anchors=titles_anchors,
-        titles_color=titles_color,
-        titles_scale=titles_scale,
-        titles_thickness=titles_thickness,
-        titles_padding=titles_padding,
-        titles_text_font=titles_text_font,
-        titles_background_color=titles_background_color,
-        default_title_placement=default_title_placement,
-    )
-    rows, columns = grid_size
-    tiles_elements = list(create_batches(sequence=images, batch_size=columns))
-    while len(tiles_elements[-1]) < columns:
-        tiles_elements[-1].append(
-            _generate_color_image(shape=single_tile_size, color=tile_padding_color)
-        )
-    while len(tiles_elements) < rows:
-        tiles_elements.append(
-            [_generate_color_image(shape=single_tile_size, color=tile_padding_color)]
-            * columns
-        )
-    return _merge_tiles_elements(
-        tiles_elements=tiles_elements,
-        grid_size=grid_size,
-        single_tile_size=single_tile_size,
-        tile_margin=tile_margin,
-        tile_margin_color=tile_margin_color,
-    )
-
-
-def _draw_texts(
-    images: list[np.ndarray],
-    titles: list[str | None] | None,
-    titles_anchors: list[Point | None],
-    titles_color: tuple[int, int, int],
-    titles_scale: float | None,
-    titles_thickness: int,
-    titles_padding: int,
-    titles_text_font: int,
-    titles_background_color: tuple[int, int, int],
-    default_title_placement: RelativePosition,
-) -> list[np.ndarray]:
-    if titles is None:
-        return images
-    titles_anchors = _prepare_default_titles_anchors(
-        images=images,
-        titles_anchors=titles_anchors,
-        default_title_placement=default_title_placement,
-    )
-    if titles_scale is None:
-        image_height, image_width = images[0].shape[:2]
-        titles_scale = calculate_optimal_text_scale(
-            resolution_wh=(image_width, image_height)
-        )
-    result = []
-    for image, text, anchor in zip(images, titles, titles_anchors):
-        if text is None:
-            result.append(image)
-            continue
-        processed_image = draw_text(
-            scene=image,
-            text=text,
-            text_anchor=anchor,
-            text_color=Color.from_bgr_tuple(titles_color),
-            text_scale=titles_scale,
-            text_thickness=titles_thickness,
-            text_padding=titles_padding,
-            text_font=titles_text_font,
-            background_color=Color.from_bgr_tuple(titles_background_color),
-        )
-        result.append(processed_image)
-    return result
-
-
-def _prepare_default_titles_anchors(
-    images: list[np.ndarray],
-    titles_anchors: list[Point | None],
-    default_title_placement: RelativePosition,
-) -> list[Point]:
-    result = []
-    for image, anchor in zip(images, titles_anchors):
-        if anchor is not None:
-            result.append(anchor)
-            continue
-        image_height, image_width = image.shape[:2]
-        if default_title_placement == "top":
-            default_anchor = Point(x=image_width / 2, y=image_height * 0.1)
-        else:
-            default_anchor = Point(x=image_width / 2, y=image_height * 0.9)
-        result.append(default_anchor)
-    return result
-
-
-def _merge_tiles_elements(
-    tiles_elements: list[list[np.ndarray]],
-    grid_size: tuple[int, int],
-    single_tile_size: tuple[int, int],
-    tile_margin: int,
-    tile_margin_color: tuple[int, int, int],
-) -> np.ndarray:
-    vertical_padding = (
-        np.ones((single_tile_size[1], tile_margin, 3)) * tile_margin_color
-    )
-    merged_rows = [
-        np.concatenate(
-            list(
-                itertools.chain.from_iterable(
-                    zip(row, [vertical_padding] * grid_size[1])
-                )
-            )[:-1],
-            axis=1,
-        )
-        for row in tiles_elements
-    ]
-    row_width = merged_rows[0].shape[1]
-    horizontal_padding = (
-        np.ones((tile_margin, row_width, 3), dtype=np.uint8) * tile_margin_color
-    )
-    rows_with_paddings = []
-    for row in merged_rows:
-        rows_with_paddings.append(row)
-        rows_with_paddings.append(horizontal_padding)
-    return np.concatenate(
-        rows_with_paddings[:-1],
-        axis=0,
-    ).astype(np.uint8)
-
-
-def _generate_color_image(
-    shape: tuple[int, int], color: tuple[int, int, int]
-) -> np.ndarray:
-    return np.ones((*shape[::-1], 3), dtype=np.uint8) * color

supervision/utils/notebook.py

@@ -4,7 +4,7 @@ import cv2
 import matplotlib.pyplot as plt
 from PIL import Image
 
-from supervision.annotators.base import ImageType
+from supervision.draw.base import ImageType
 from supervision.utils.conversion import pillow_to_cv2
 
 

supervision/validators/__init__.py

@@ -53,7 +53,7 @@ def validate_confidence(confidence: Any, n: int) -> None:
         )
 
 
-def validate_keypoint_confidence(confidence: Any, n: int, m: int) -> None:
+def validate_key_point_confidence(confidence: Any, n: int, m: int) -> None:
     expected_shape = f"({n, m})"
     actual_shape = str(getattr(confidence, "shape", None))
 
@@ -126,7 +126,7 @@ def validate_detections_fields(
     validate_data(data, n)
 
 
-def validate_keypoints_fields(
+def validate_key_points_fields(
     xy: Any,
     class_id: Any,
     confidence: Any,
@@ -136,7 +136,7 @@ def validate_keypoints_fields(
     m = len(xy[0]) if len(xy) > 0 else 0
     validate_xy(xy, n, m)
     validate_class_id(class_id, n)
-    validate_keypoint_confidence(confidence, n, m)
+    validate_key_point_confidence(confidence, n, m)
     validate_data(data, n)
 
 

test/annotators/test_utils.py

@@ -5,7 +5,7 @@ from contextlib import ExitStack as DoesNotRaise
 import numpy as np
 import pytest
 
-from supervision.annotators.utils import ColorLookup, resolve_color_idx
+from supervision.annotators.utils import ColorLookup, resolve_color_idx, wrap_text
 from supervision.detection.core import Detections
 from test.test_utils import mock_detections
 
@@ -97,7 +97,7 @@ from test.test_utils import mock_detections
 def test_resolve_color_idx(
     detections: Detections,
     detection_idx: int,
-    color_lookup: ColorLookup,
+    color_lookup: ColorLookup | np.ndarray,
     expected_result: int | None,
     exception: Exception,
 ) -> None:
@@ -108,3 +108,67 @@ def test_resolve_color_idx(
             color_lookup=color_lookup,
         )
         assert result == expected_result
+
+
+@pytest.mark.parametrize(
+    "text, max_line_length, expected_result, exception",
+    [
+        (None, None, [""], DoesNotRaise()),  # text is None
+        ("", None, [""], DoesNotRaise()),  # empty string
+        ("   \t  ", 3, [""], DoesNotRaise()),  # whitespace-only (spaces + tab)
+        (12345, None, ["12345"], DoesNotRaise()),  # plain integer
+        (-6789, None, ["-6789"], DoesNotRaise()),  # negative integer
+        (np.int64(1000), None, ["1000"], DoesNotRaise()),  # NumPy int64
+        ([1, 2, 3], None, ["[1, 2, 3]"], DoesNotRaise()),  # list to string
+        (
+            "When you play the game of thrones, you win or you die.\nFear cuts deeper than swords.\nA mind needs books as a sword needs a whetstone.",  # noqa: E501
+            None,
+            [
+                "When you play the game of thrones, you win or you die.",
+                "Fear cuts deeper than swords.",
+                "A mind needs books as a sword needs a whetstone.",
+            ],
+            DoesNotRaise(),
+        ),  # Game-of-Thrones quotes, multiline
+        ("\n", None, [""], DoesNotRaise()),  # single newline
+        (
+            "valarmorghulisvalardoharis",
+            6,
+            ["valarm", "orghul", "isvala", "rdohar", "is"],
+            DoesNotRaise(),
+        ),  # long Valyrian phrase, wrapped
+        (
+            "Winter is coming\nFire and blood",
+            10,
+            [
+                "Winter is",
+                "coming",
+                "Fire and",
+                "blood",
+            ],
+            DoesNotRaise(),
+        ),  # mix of short/long with newline
+        (
+            "What is dead may never die",
+            0,
+            None,
+            pytest.raises(ValueError),
+        ),  # width 0 - invalid
+        (
+            "A Lannister always pays his debts",
+            -1,
+            None,
+            pytest.raises(ValueError),
+        ),  # width -1 - invalid
+        (None, 10, [""], DoesNotRaise()),  # text None, width set
+    ],
+)
+def test_wrap_text(
+    text: object,
+    max_line_length: int | None,
+    expected_result: list[str],
+    exception: Exception,
+) -> None:
+    with exception:
+        result = wrap_text(text=text, max_line_length=max_line_length)
+        assert result == expected_result

test/detection/test_vlm.py

@@ -6,7 +6,10 @@ from contextlib import nullcontext as does_not_raise
 import numpy as np
 import pytest
 
+from supervision.config import CLASS_NAME_DATA_FIELD
+from supervision.detection.core import Detections
 from supervision.detection.vlm import (
+    VLM,
     from_florence_2,
     from_google_gemini_2_0,
     from_google_gemini_2_5,
@@ -1122,3 +1125,110 @@ def test_from_google_gemini_2_5(
             assert masks is not None
             assert masks.shape == expected_results[4].shape
             assert np.array_equal(masks, expected_results[4])
+
+
+@pytest.mark.parametrize(
+    "exception, result, resolution_wh, classes, expected_detections",
+    [
+        (
+            pytest.raises(ValueError),
+            "",
+            (100, 100),
+            None,
+            None,
+        ),  # empty text
+        (
+            pytest.raises(ValueError),
+            "random text",
+            (100, 100),
+            None,
+            None,
+        ),  # random text
+        (
+            does_not_raise(),
+            "<|ref|>cat<|/ref|><|det|>[[100, 200, 300, 400]]<|/det|>",
+            (1000, 1000),
+            None,
+            Detections(
+                xyxy=np.array([[100.1, 200.2, 300.3, 400.4]]),
+                class_id=np.array([0]),
+                data={CLASS_NAME_DATA_FIELD: np.array(["cat"])},
+            ),
+        ),  # single box, no classes
+        (
+            does_not_raise(),
+            "<|ref|>cat<|/ref|><|det|>[[100, 200, 300, 400]]<|/det|>",
+            (1000, 1000),
+            ["cat", "dog"],
+            Detections(
+                xyxy=np.array([[100.1, 200.2, 300.3, 400.4]]),
+                class_id=np.array([0]),
+                data={CLASS_NAME_DATA_FIELD: np.array(["cat"])},
+            ),
+        ),  # single box, with classes
+        (
+            does_not_raise(),
+            "<|ref|>person<|/ref|><|det|>[[100, 200, 300, 400]]<|/det|>",
+            (1000, 1000),
+            ["cat", "dog"],
+            Detections.empty(),
+        ),  # single box, wrong class
+        (
+            does_not_raise(),
+            (
+                "<|ref|>cat<|/ref|><|det|>[[100, 200, 300, 400]]<|/det|>"
+                "<|ref|>dog<|/ref|><|det|>[[500, 600, 700, 800]]<|/det|>"
+            ),
+            (1000, 1000),
+            ["cat"],
+            Detections(
+                xyxy=np.array([[100.1, 200.2, 300.3, 400.4]]),
+                class_id=np.array([0]),
+                data={CLASS_NAME_DATA_FIELD: np.array(["cat"])},
+            ),
+        ),  # multiple boxes, one class correct
+        (
+            pytest.raises(ValueError),
+            "<|ref|>cat<|/ref|>",
+            (100, 100),
+            None,
+            None,
+        ),  # only ref
+        (
+            pytest.raises(ValueError),
+            "<|det|>[[100, 200, 300, 400]]<|/det|>",
+            (100, 100),
+            None,
+            None,
+        ),  # only det
+    ],
+)
+def test_from_deepseek_vl_2(
+    exception,
+    result: str,
+    resolution_wh: tuple[int, int],
+    classes: list[str] | None,
+    expected_detections: Detections,
+):
+    with exception:
+        detections = Detections.from_vlm(
+            vlm=VLM.DEEPSEEK_VL_2,
+            result=result,
+            resolution_wh=resolution_wh,
+            classes=classes,
+        )
+
+        if expected_detections is None:
+            return
+
+        assert len(detections) == len(expected_detections)
+
+        if len(detections) == 0:
+            return
+
+        assert np.allclose(detections.xyxy, expected_detections.xyxy, atol=1e-1)
+        assert np.array_equal(detections.class_id, expected_detections.class_id)
+        assert np.array_equal(
+            detections.data[CLASS_NAME_DATA_FIELD],
+            expected_detections.data[CLASS_NAME_DATA_FIELD],
+        )

test/detection/utils/test_converters.py

@@ -6,6 +6,7 @@ import pytest
 from supervision.detection.utils.converters import (
     xcycwh_to_xyxy,
     xywh_to_xyxy,
+    xyxy_to_mask,
     xyxy_to_xcycarh,
     xyxy_to_xywh,
 )
@@ -129,3 +130,174 @@ def test_xyxy_to_xcycarh(xyxy: np.ndarray, expected_result: np.ndarray) -> None:
 def test_xcycwh_to_xyxy(xcycwh: np.ndarray, expected_result: np.ndarray) -> None:
     result = xcycwh_to_xyxy(xcycwh)
     np.testing.assert_array_equal(result, expected_result)
+
+
+@pytest.mark.parametrize(
+    "boxes,resolution_wh,expected",
+    [
+        # 0) Empty input
+        (
+            np.array([], dtype=float).reshape(0, 4),
+            (5, 4),
+            np.array([], dtype=bool).reshape(0, 4, 5),
+        ),
+        # 1) Single pixel box
+        (
+            np.array([[2, 1, 2, 1]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [False, False, False, False, False],
+                        [False, False, True, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 2) Horizontal line, inclusive bounds
+        (
+            np.array([[1, 2, 3, 2]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                        [False, True, True, True, False],
+                        [False, False, False, False, False],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 3) Vertical line, inclusive bounds
+        (
+            np.array([[3, 0, 3, 2]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [False, False, False, True, False],
+                        [False, False, False, True, False],
+                        [False, False, False, True, False],
+                        [False, False, False, False, False],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 4) Proper rectangle fill
+        (
+            np.array([[1, 1, 3, 2]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [False, False, False, False, False],
+                        [False, True, True, True, False],
+                        [False, True, True, True, False],
+                        [False, False, False, False, False],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 5) Negative coordinates clipped to [0, 0]
+        (
+            np.array([[-2, -1, 1, 1]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [True, True, False, False, False],
+                        [True, True, False, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 6) Overflow coordinates clipped to width-1 and height-1
+        (
+            np.array([[3, 2, 10, 10]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, True, True],
+                        [False, False, False, True, True],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 7) Invalid box where max < min after ints, mask stays empty
+        (
+            np.array([[3, 2, 1, 4]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 8) Fractional coordinates are floored by int conversion
+        #    (0.2,0.2)-(2.8,1.9) -> (0,0)-(2,1)
+        (
+            np.array([[0.2, 0.2, 2.8, 1.9]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    [
+                        [True, True, True, False, False],
+                        [True, True, True, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                    ]
+                ],
+                dtype=bool,
+            ),
+        ),
+        # 9) Multiple boxes, separate masks
+        (
+            np.array([[0, 0, 1, 0], [2, 1, 4, 3]], dtype=float),
+            (5, 4),
+            np.array(
+                [
+                    # Box 0: row 0, cols 0..1
+                    [
+                        [True, True, False, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                        [False, False, False, False, False],
+                    ],
+                    # Box 1: rows 1..3, cols 2..4
+                    [
+                        [False, False, False, False, False],
+                        [False, False, True, True, True],
+                        [False, False, True, True, True],
+                        [False, False, True, True, True],
+                    ],
+                ],
+                dtype=bool,
+            ),
+        ),
+    ],
+)
+def test_xyxy_to_mask(boxes: np.ndarray, resolution_wh, expected: np.ndarray) -> None:
+    result = xyxy_to_mask(boxes, resolution_wh)
+    assert result.dtype == np.bool_
+    assert result.shape == expected.shape
+    np.testing.assert_array_equal(result, expected)

test/detection/utils/test_vlms.py

@@ -0,0 +1,117 @@
+import pytest
+
+from supervision.detection.utils.vlms import edit_distance, fuzzy_match_index
+
+
+@pytest.mark.parametrize(
+    "string_1, string_2, case_sensitive, expected_result",
+    [
+        # identical strings, various cases
+        ("hello", "hello", True, 0),
+        ("hello", "hello", False, 0),
+        # case sensitive vs insensitive
+        ("Test", "test", True, 1),
+        ("Test", "test", False, 0),
+        ("CASE", "case", True, 4),
+        ("CASE", "case", False, 0),
+        # completely different
+        ("abc", "xyz", True, 3),
+        ("abc", "xyz", False, 3),
+        # one string empty
+        ("hello", "", True, 5),
+        ("", "world", True, 5),
+        # single character cases
+        ("a", "b", True, 1),
+        ("A", "a", True, 1),
+        ("A", "a", False, 0),
+        # whitespaces
+        ("hello world", "helloworld", True, 1),
+        ("test", " test", True, 1),
+        # unicode and emoji
+        ("😊", "😊", True, 0),
+        ("😊", "😢", True, 1),
+        # long string vs empty
+        ("a" * 100, "", True, 100),
+        ("", "b" * 100, True, 100),
+        # prefix/suffix
+        ("prefix", "prefixes", True, 2),
+        ("suffix", "asuffix", True, 1),
+        # leading/trailing whitespace
+        (" hello", "hello", True, 1),
+        ("hello", "hello ", True, 1),
+        # long almost-equal string
+        (
+            "The quick brown fox jumps over the lazy dog",
+            "The quick brown fox jumps over the lazy cog",
+            True,
+            1,
+        ),
+        (
+            "The quick brown fox jumps over the lazy dog",
+            "The quick brown fox jumps over the lazy cog",
+            False,
+            1,
+        ),
+        # both empty
+        ("", "", True, 0),
+        ("", "", False, 0),
+        # mixed case with symbols
+        ("123ABC!", "123abc!", True, 3),
+        ("123ABC!", "123abc!", False, 0),
+    ],
+)
+def test_edit_distance(string_1, string_2, case_sensitive, expected_result):
+    assert (
+        edit_distance(string_1, string_2, case_sensitive=case_sensitive)
+        == expected_result
+    )
+
+
+@pytest.mark.parametrize(
+    "candidates, query, threshold, case_sensitive, expected_result",
+    [
+        # exact match at index 0
+        (["cat", "dog", "rat"], "cat", 0, True, 0),
+        # match at index 2 within threshold
+        (["cat", "dog", "rat"], "dat", 1, True, 0),
+        # no match due to high threshold
+        (["cat", "dog", "rat"], "bat", 0, True, None),
+        # multiple possible matches, returns first
+        (["apple", "apply", "appla"], "apple", 1, True, 0),
+        # case-insensitive match
+        (["Alpha", "beta", "Gamma"], "alpha", 0, False, 0),
+        # case-sensitive: no match
+        (["Alpha", "beta", "Gamma"], "alpha", 0, True, None),
+        # threshold boundary
+        (["alpha", "beta", "gamma"], "bata", 1, True, 1),
+        # no match (all distances too high)
+        (["one", "two", "three"], "ten", 1, True, None),
+        # unicode/emoji match
+        (["😊", "😢", "😁"], "😄", 1, True, 0),
+        (["😊", "😢", "😁"], "😊", 0, True, 0),
+        # empty candidates
+        ([], "any", 2, True, None),
+        # empty query, non-empty candidates
+        (["", "abc"], "", 0, True, 0),
+        (["", "abc"], "", 1, True, 0),
+        (["a", "b", "c"], "", 1, True, 0),
+        # non-empty query, empty candidate
+        (["", ""], "a", 1, True, 0),
+        # all candidates require higher edit than threshold
+        (["short", "words", "only"], "longerword", 2, True, None),
+        # repeated candidates
+        (["a", "a", "a"], "b", 1, True, 0),
+    ],
+)
+def test_fuzzy_match_index(
+    candidates, query, threshold, case_sensitive, expected_result
+):
+    assert (
+        fuzzy_match_index(
+            candidates=candidates,
+            query=query,
+            threshold=threshold,
+            case_sensitive=case_sensitive,
+        )
+        == expected_result
+    )

test/key_points/test_core.py

@@ -0,0 +1,268 @@
+from contextlib import nullcontext as DoesNotRaise
+
+import numpy as np
+import pytest
+
+from supervision.key_points.core import KeyPoints
+from test.test_utils import mock_key_points
+
+KEY_POINTS = mock_key_points(
+    xy=[
+        [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]],
+        [[10, 11], [12, 13], [14, 15], [16, 17], [18, 19]],
+        [[20, 21], [22, 23], [24, 25], [26, 27], [28, 29]],
+    ],
+    confidence=[
+        [0.8, 0.2, 0.6, 0.1, 0.5],
+        [0.7, 0.9, 0.3, 0.4, 0.0],
+        [0.1, 0.6, 0.8, 0.2, 0.7],
+    ],
+    class_id=[0, 1, 2],
+)
+
+
+@pytest.mark.parametrize(
+    "key_points, index, expected_result, exception",
+    [
+        (
+            KeyPoints.empty(),
+            slice(None),
+            KeyPoints.empty(),
+            DoesNotRaise(),
+        ),  # slice all key points when key points object empty
+        (
+            KEY_POINTS,
+            slice(None),
+            KEY_POINTS,
+            DoesNotRaise(),
+        ),  # slice all key points when key points object nonempty
+        (
+            KEY_POINTS,
+            slice(0, 1),
+            mock_key_points(
+                xy=[[[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]]],
+                confidence=[[0.8, 0.2, 0.6, 0.1, 0.5]],
+                class_id=[0],
+            ),
+            DoesNotRaise(),
+        ),  # select the first skeleton by slice
+        (
+            KEY_POINTS,
+            slice(0, 2),
+            mock_key_points(
+                xy=[
+                    [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]],
+                    [[10, 11], [12, 13], [14, 15], [16, 17], [18, 19]],
+                ],
+                confidence=[
+                    [0.8, 0.2, 0.6, 0.1, 0.5],
+                    [0.7, 0.9, 0.3, 0.4, 0.0],
+                ],
+                class_id=[0, 1],
+            ),
+            DoesNotRaise(),
+        ),  # select the first skeleton by slice
+        (
+            KEY_POINTS,
+            0,
+            mock_key_points(
+                xy=[[[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]]],
+                confidence=[[0.8, 0.2, 0.6, 0.1, 0.5]],
+                class_id=[0],
+            ),
+            DoesNotRaise(),
+        ),  # select the first skeleton by index
+        (
+            KEY_POINTS,
+            -1,
+            mock_key_points(
+                xy=[[[20, 21], [22, 23], [24, 25], [26, 27], [28, 29]]],
+                confidence=[[0.1, 0.6, 0.8, 0.2, 0.7]],
+                class_id=[2],
+            ),
+            DoesNotRaise(),
+        ),  # select the last skeleton by index
+        (
+            KEY_POINTS,
+            [0, 1],
+            mock_key_points(
+                xy=[
+                    [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]],
+                    [[10, 11], [12, 13], [14, 15], [16, 17], [18, 19]],
+                ],
+                confidence=[
+                    [0.8, 0.2, 0.6, 0.1, 0.5],
+                    [0.7, 0.9, 0.3, 0.4, 0.0],
+                ],
+                class_id=[0, 1],
+            ),
+            DoesNotRaise(),
+        ),  # select the first two skeletons by index; list
+        (
+            KEY_POINTS,
+            np.array([0, 1]),
+            mock_key_points(
+                xy=[
+                    [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]],
+                    [[10, 11], [12, 13], [14, 15], [16, 17], [18, 19]],
+                ],
+                confidence=[
+                    [0.8, 0.2, 0.6, 0.1, 0.5],
+                    [0.7, 0.9, 0.3, 0.4, 0.0],
+                ],
+                class_id=[0, 1],
+            ),
+            DoesNotRaise(),
+        ),  # select the first two skeletons by index; np.array
+        (
+            KEY_POINTS,
+            [True, True, False],
+            mock_key_points(
+                xy=[
+                    [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]],
+                    [[10, 11], [12, 13], [14, 15], [16, 17], [18, 19]],
+                ],
+                confidence=[
+                    [0.8, 0.2, 0.6, 0.1, 0.5],
+                    [0.7, 0.9, 0.3, 0.4, 0.0],
+                ],
+                class_id=[0, 1],
+            ),
+            DoesNotRaise(),
+        ),  # select only skeletons associated with positive filter; list
+        (
+            KEY_POINTS,
+            np.array([True, True, False]),
+            mock_key_points(
+                xy=[
+                    [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]],
+                    [[10, 11], [12, 13], [14, 15], [16, 17], [18, 19]],
+                ],
+                confidence=[
+                    [0.8, 0.2, 0.6, 0.1, 0.5],
+                    [0.7, 0.9, 0.3, 0.4, 0.0],
+                ],
+                class_id=[0, 1],
+            ),
+            DoesNotRaise(),
+        ),  # select only skeletons associated with positive filter; list
+        (
+            KEY_POINTS,
+            (slice(None), slice(None)),
+            KEY_POINTS,
+            DoesNotRaise(),
+        ),  # slice all anchors from all skeletons
+        (
+            KEY_POINTS,
+            (slice(None), slice(0, 1)),
+            mock_key_points(
+                xy=[[[0, 1]], [[10, 11]], [[20, 21]]],
+                confidence=[[0.8], [0.7], [0.1]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # slice the first anchor from every skeleton
+        (
+            KEY_POINTS,
+            (slice(None), slice(0, 2)),
+            mock_key_points(
+                xy=[[[0, 1], [2, 3]], [[10, 11], [12, 13]], [[20, 21], [22, 23]]],
+                confidence=[[0.8, 0.2], [0.7, 0.9], [0.1, 0.6]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # slice the first anchor two anchors from every skeleton
+        (
+            KEY_POINTS,
+            (slice(None), 0),
+            mock_key_points(
+                xy=[[[0, 1]], [[10, 11]], [[20, 21]]],
+                confidence=[[0.8], [0.7], [0.1]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # select the first anchor from every skeleton by index
+        (
+            KEY_POINTS,
+            (slice(None), -1),
+            mock_key_points(
+                xy=[[[8, 9]], [[18, 19]], [[28, 29]]],
+                confidence=[[0.5], [0.0], [0.7]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # select the last anchor from every skeleton by index
+        (
+            KEY_POINTS,
+            (slice(None), [0, 1]),
+            mock_key_points(
+                xy=[[[0, 1], [2, 3]], [[10, 11], [12, 13]], [[20, 21], [22, 23]]],
+                confidence=[[0.8, 0.2], [0.7, 0.9], [0.1, 0.6]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # select the first two anchors from every skeleton by index; list
+        (
+            KEY_POINTS,
+            (slice(None), np.array([0, 1])),
+            mock_key_points(
+                xy=[[[0, 1], [2, 3]], [[10, 11], [12, 13]], [[20, 21], [22, 23]]],
+                confidence=[[0.8, 0.2], [0.7, 0.9], [0.1, 0.6]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # select the first two anchors from every skeleton by index; np.array
+        (
+            KEY_POINTS,
+            (slice(None), [True, True, False, False, False]),
+            mock_key_points(
+                xy=[[[0, 1], [2, 3]], [[10, 11], [12, 13]], [[20, 21], [22, 23]]],
+                confidence=[[0.8, 0.2], [0.7, 0.9], [0.1, 0.6]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # select only anchors associated with positive filter; list
+        (
+            KEY_POINTS,
+            (slice(None), np.array([True, True, False, False, False])),
+            mock_key_points(
+                xy=[[[0, 1], [2, 3]], [[10, 11], [12, 13]], [[20, 21], [22, 23]]],
+                confidence=[[0.8, 0.2], [0.7, 0.9], [0.1, 0.6]],
+                class_id=[0, 1, 2],
+            ),
+            DoesNotRaise(),
+        ),  # select only anchors associated with positive filter; np.array
+        (
+            KEY_POINTS,
+            (0, 0),
+            mock_key_points(
+                xy=[
+                    [[0, 1]],
+                ],
+                confidence=[
+                    [0.8],
+                ],
+                class_id=[0],
+            ),
+            DoesNotRaise(),
+        ),  # select the first anchor from the first skeleton by index
+        (
+            KEY_POINTS,
+            (0, -1),
+            mock_key_points(
+                xy=[
+                    [[8, 9]],
+                ],
+                confidence=[
+                    [0.5],
+                ],
+                class_id=[0],
+            ),
+            DoesNotRaise(),
+        ),  # select the last anchor from the first skeleton by index
+    ],
+)
+def test_key_points_getitem(key_points, index, expected_result, exception):
+    with exception:
+        result = key_points[index]
+        assert result == expected_result

test/test_utils.py

@@ -3,14 +3,13 @@ from __future__ import annotations
 from typing import Any
 
 import numpy as np
-import numpy.typing as npt
 
 from supervision.detection.core import Detections
-from supervision.keypoint.core import KeyPoints
+from supervision.key_points.core import KeyPoints
 
 
 def mock_detections(
-    xyxy: npt.NDArray[np.float32],
+    xyxy: list[list[float]],
     mask: list[np.ndarray] | None = None,
     confidence: list[float] | None = None,
     class_id: list[int] | None = None,
@@ -34,9 +33,9 @@ def mock_detections(
     )
 
 
-def mock_keypoints(
-    xy: npt.NDArray[np.float32],
-    confidence: list[float] | None = None,
+def mock_key_points(
+    xy: list[list[list[float]]],
+    confidence: list[list[float]] | None = None,
     class_id: list[int] | None = None,
     data: dict[str, list[Any]] | None = None,
 ) -> KeyPoints:

test/utils/test_conversion.py

@@ -3,7 +3,7 @@ from PIL import Image, ImageChops
 
 from supervision.utils.conversion import (
     cv2_to_pillow,
-    ensure_cv2_image_for_processing,
+    ensure_cv2_image_for_standalone_function,
     images_to_cv2,
     pillow_to_cv2,
 )
@@ -16,7 +16,7 @@ def test_ensure_cv2_image_for_processing_when_pillow_image_submitted(
     param_a_value = 3
     param_b_value = "some"
 
-    @ensure_cv2_image_for_processing
+    @ensure_cv2_image_for_standalone_function
     def my_custom_processing_function(
         image: np.ndarray,
         param_a: int,
@@ -55,7 +55,7 @@ def test_ensure_cv2_image_for_processing_when_cv2_image_submitted(
     param_a_value = 3
     param_b_value = "some"
 
-    @ensure_cv2_image_for_processing
+    @ensure_cv2_image_for_standalone_function
     def my_custom_processing_function(
         image: np.ndarray,
         param_a: int,

test/utils/test_image.py

@@ -1,9 +1,7 @@
 import numpy as np
-import pytest
 from PIL import Image, ImageChops
 
-from supervision import Color, Point
-from supervision.utils.image import create_tiles, letterbox_image, resize_image
+from supervision.utils.image import letterbox_image, resize_image
 
 
 def test_resize_image_for_opencv_image() -> None:
@@ -96,147 +94,3 @@ def test_letterbox_image_for_pillow_image() -> None:
     assert difference.getbbox() is None, (
         "Expected padding to be added top and bottom with padding added top and bottom"
     )
-
-
-def test_create_tiles_with_one_image(
-    one_image: np.ndarray, single_image_tile: np.ndarray
-) -> None:
-    # when
-    result = create_tiles(images=[one_image], single_tile_size=(240, 240))
-
-    # # then
-    assert np.allclose(result, single_image_tile, atol=5.0)
-
-
-def test_create_tiles_with_one_image_and_enforced_grid(
-    one_image: np.ndarray, single_image_tile_enforced_grid: np.ndarray
-) -> None:
-    # when
-    result = create_tiles(
-        images=[one_image],
-        grid_size=(None, 3),
-        single_tile_size=(240, 240),
-    )
-
-    # then
-    assert np.allclose(result, single_image_tile_enforced_grid, atol=5.0)
-
-
-def test_create_tiles_with_two_images(
-    two_images: list[np.ndarray], two_images_tile: np.ndarray
-) -> None:
-    # when
-    result = create_tiles(images=two_images, single_tile_size=(240, 240))
-
-    # then
-    assert np.allclose(result, two_images_tile, atol=5.0)
-
-
-def test_create_tiles_with_three_images(
-    three_images: list[np.ndarray], three_images_tile: np.ndarray
-) -> None:
-    # when
-    result = create_tiles(images=three_images, single_tile_size=(240, 240))
-
-    # then
-    assert np.allclose(result, three_images_tile, atol=5.0)
-
-
-def test_create_tiles_with_four_images(
-    four_images: list[np.ndarray],
-    four_images_tile: np.ndarray,
-) -> None:
-    # when
-    result = create_tiles(images=four_images, single_tile_size=(240, 240))
-
-    # then
-    assert np.allclose(result, four_images_tile, atol=5.0)
-
-
-def test_create_tiles_with_all_images(
-    all_images: list[np.ndarray],
-    all_images_tile: np.ndarray,
-) -> None:
-    # when
-    result = create_tiles(images=all_images, single_tile_size=(240, 240))
-
-    # then
-    assert np.allclose(result, all_images_tile, atol=5.0)
-
-
-def test_create_tiles_with_all_images_and_custom_grid(
-    all_images: list[np.ndarray], all_images_tile_and_custom_grid: np.ndarray
-) -> None:
-    # when
-    result = create_tiles(
-        images=all_images,
-        grid_size=(3, 3),
-        single_tile_size=(240, 240),
-    )
-
-    # then
-    assert np.allclose(result, all_images_tile_and_custom_grid, atol=5.0)
-
-
-def test_create_tiles_with_all_images_and_custom_colors(
-    all_images: list[np.ndarray], all_images_tile_and_custom_colors: np.ndarray
-) -> None:
-    # when
-    result = create_tiles(
-        images=all_images,
-        tile_margin_color=(127, 127, 127),
-        tile_padding_color=(224, 224, 224),
-        single_tile_size=(240, 240),
-    )
-
-    # then
-    assert np.allclose(result, all_images_tile_and_custom_colors, atol=5.0)
-
-
-def test_create_tiles_with_all_images_and_titles(
-    all_images: list[np.ndarray],
-    all_images_tile_and_custom_colors_and_titles: np.ndarray,
-) -> None:
-    # when
-    result = create_tiles(
-        images=all_images,
-        titles=["Image 1", None, "Image 3", "Image 4"],
-        single_tile_size=(240, 240),
-    )
-
-    # then
-    assert np.allclose(result, all_images_tile_and_custom_colors_and_titles, atol=5.0)
-
-
-def test_create_tiles_with_all_images_and_titles_with_custom_configs(
-    all_images: list[np.ndarray],
-    all_images_tile_and_titles_with_custom_configs: np.ndarray,
-) -> None:
-    # when
-    result = create_tiles(
-        images=all_images,
-        titles=["Image 1", None, "Image 3", "Image 4"],
-        single_tile_size=(240, 240),
-        titles_anchors=[
-            Point(x=200, y=300),
-            Point(x=300, y=400),
-            None,
-            Point(x=300, y=400),
-        ],
-        titles_color=Color.RED,
-        titles_scale=1.5,
-        titles_thickness=3,
-        titles_padding=20,
-        titles_background_color=Color.BLACK,
-        default_title_placement="bottom",
-    )
-
-    # then
-    assert np.allclose(result, all_images_tile_and_titles_with_custom_configs, atol=5.0)
-
-
-def test_create_tiles_with_all_images_and_custom_grid_to_small_to_fit_images(
-    all_images: list[np.ndarray],
-) -> None:
-    with pytest.raises(ValueError):
-        _ = create_tiles(images=all_images, grid_size=(2, 2))