fix(dataset): export annotations for in-memory grayscale images - #2641

Description

Exporting a DetectionDataset containing a two-dimensional grayscale image raises a dimension-unpacking error in LabelMe and YOLO, or an IndexError in Pascal VOC. Read height and width independently of channels, and write Pascal VOC depth as 1 for grayscale images.

Minimal reproduction on develop at 8c7df2388ce46f16139edc498c2bc3874493b4bc:

import numpy as np
import supervision as sv

dataset = sv.DetectionDataset(
    classes=["object"],
    images={"gray.png": np.zeros((16, 24), dtype=np.uint8)},
    annotations={
        "gray.png": sv.Detections(
            xyxy=np.array([[3, 2, 15, 10]], dtype=np.float32),
            class_id=np.array([0]),
        )
    },
)
dataset.as_labelme(annotations_directory_path="labels")
# ValueError: not enough values to unpack (expected 3, got 2)
# as_yolo fails the same way; as_pascal_voc raises IndexError.

Motivation and Context

In-memory image dictionaries remain supported until their scheduled removal in 0.33.0. COCO export already accepts these grayscale arrays. The other three exporters should preserve their dimensions and annotations too.

PR #2410 addressed image modes during loading. PR #2637, currently open, addresses copying images into their source folder. Neither fixes grayscale annotation dimensions.

Changes Made

• Read the first two shape dimensions in LabelMe and YOLO; document the supported YOLO shape tuples.
• Set Pascal VOC depth to 1 for two-dimensional images without changing image pixels.
• Add 12 format-local regression cases covering grayscale/color images, boxes/masks, dimensions, coordinates, and unchanged source arrays.
• Add an Unreleased changelog entry. The in-memory API's deprecation schedule remains unchanged.

Testing

On unmodified production code, the new tests failed for all six grayscale cases with the expected dimension errors; all six color controls passed:

uv run pytest tests/dataset/formats/test_labelme.py tests/dataset/formats/test_yolo.py tests/dataset/formats/test_pascal_voc.py -k ImageShapes
6 failed, 6 passed, 161 deselected

With the fix:

uv run pytest tests/dataset/formats/test_labelme.py tests/dataset/formats/test_yolo.py tests/dataset/formats/test_pascal_voc.py
173 passed

uv run pytest --cov=supervision
4329 passed, 24 skipped; 91% coverage

uv run pre-commit run --all-files
All hooks passed

uv run pytest --doctest-modules src/supervision/dataset/formats/labelme.py src/supervision/dataset/formats/yolo.py src/supervision/dataset/formats/pascal_voc.py
5 passed

The clean base passed 4316 tests with the same 24 skips and 91% coverage; its pre-commit checks also passed. Validation ran on macOS with Python 3.13.9 using the NumPy fallback backend. Native OpenCV and rasterio were not installed; their existing optional reference tests account for the skips. No test skips were added.

• Added/updated tests, and the full suite passes locally
• Updated docs (docstrings / mkdocs entry) for new or changed public API
• Added a changelog entry in docs/changelog.md under Unreleased (skip for lint/type/format-only or pure doc changes)

Additional Notes

No public API, dependencies, coordinate conventions, or image pixels are changed.