[Pipeline] Add top_k, label filtering, box_format and score sorting to ObjectDetectionPipeline

[vimal-crypto]

What this PR does

This PR brings ObjectDetectionPipeline in line with its sister pipelines (ZeroShotObjectDetectionPipeline, ImageClassificationPipeline) by adding four enhancements to the postprocessing stage.

Changes

1. Score sorting — results are now always returned sorted by descending confidence score, consistent with ZeroShotObjectDetectionPipeline and ImageClassificationPipeline. Previously ObjectDetectionPipeline was the only vision pipeline that returned results in raw model-output order (anchor order), which was unintuitive and undocumented.

2. top_k parameter — allows users to cap the number of returned detections to the N highest-confidence results. Defaults to None (return all detections above threshold), preserving full backward compatibility.

3. labels parameter — accepts a list of class-name strings; only detections whose label appears in the list are returned. This is a novel filtering capability with no prior equivalent in the standard object detection pipeline. Defaults to None (return all detected classes).

4. box_format parameter — controls the coordinate format of the returned bounding boxes. Supported values:

• "xyxy" (default, backward-compatible): {xmin, ymin, xmax, ymax} in pixels
• "xywh": {x_center, y_center, width, height} in pixels
• "normalized": {xmin, ymin, xmax, ymax} as floats in [0, 1] relative to image dimensions

The _get_bounding_box helper is extended to handle all three formats in one place.

Motivation

• ZeroShotObjectDetectionPipeline already has top_k and sorted outputs (see postprocess)
• ImageClassificationPipeline already has top_k and sorted outputs
• ObjectDetectionPipeline was the only vision pipeline missing these features, creating an inconsistency across the pipeline family
• box_format and labels are new capabilities addressing common downstream use-cases (COCO eval tools, Roboflow, OpenCV, torchvision draw utilities all expect different box formats; class filtering is a very common need)

Breaking changes

None. All four new parameters are fully optional with safe defaults:

• top_k=None → no truncation
• labels=None → no filtering
• box_format="xyxy" → identical output to before
• Score sorting is the only behavioral change to default output; results are now consistently ordered highest-to-lowest (expected order by virtually all downstream consumers)

Existing slow test expected outputs in test_large_model_pt and test_integration_torch_object_detection are updated to reflect the new sorted order.

Tests added

8 new unit tests covering all new parameters and edge cases:

Test	Enhancement	Speed
test_top_k	top_k truncation to N results	Fast
test_results_sorted_by_score	score sort order guarantee	Fast
test_label_filter	label allowlist filtering (match)	Slow
test_label_filter_excludes_all	label allowlist (no match → empty list)	Fast
test_box_format_xyxy	xyxy format returns int keys	Fast
test_box_format_xywh	xywh format returns correct keys and positive dims	Fast
test_box_format_normalized	normalized format returns floats in [0,1]	Fast
test_box_format_invalid_raises	unsupported format raises ValueError	Fast

Files changed

• src/transformers/pipelines/object_detection.py
• tests/pipelines/test_pipelines_object_detection.py