fix(detectnet): truncate clustered boxes to MAX_BOXES to avoid broadc…

[andrewwhitecdw]

…ast crash

Summary

cluster() in the DetectNet Python layers (ClusterGroundtruth / ClusterDetections) converts grid-format coverage/bbox network output into a fixed-size [batch_size, MAX_BOXES, 5] blob. When an image produces more than MAX_BOXES (50) box proposals, the forward pass dies with:

ValueError: could not broadcast input array from shape (256,4) into shape (50,4)

so any test/val image with >50 ground-truth objects (or >50 clustered detections) aborts the whole test phase instead of just keeping the 50 slots the blob has room for.

Root cause

boxes = np.zeros([batch_size, MAX_BOXES, 5])
...
[r, c] = boxes_cur_image.shape
boxes[i, 0:r, 0:c] = boxes_cur_image

boxes[i] only has MAX_BOXES rows, but r (the number of proposals that survived thresholding / dedup / groupRectangles voting) is unbounded, so the slice assignment fails to broadcast whenever r > MAX_BOXES.

Repro: 16x16 grid, stride 1, all 256 cells covered => 256 ground-truth proposals => ValueError: could not broadcast input array from shape (256,4) into shape (50,4).

Fix

r = min(r, MAX_BOXES)
boxes[i, 0:r, 0:c] = boxes_cur_image[0:r]

Clip the copy to the blob capacity, which is exactly what the "max_bbox_per_image = MAX_BOXES" contract documented in the layer docstrings implies.

Testing

No GPU/caffe build needed: loaded clustering.py standalone with caffe and cv2 stubbed out and drove cluster() with a synthetic batch of 256 covered grid cells (uv run --with numpy python repro.py):

• pre-fix (stash): ValueError: could not broadcast input array from shape (256,4) into shape (50,4) — crash reproduced.
• post-fix: passes, output shape (1, 50, 5) with all 50 slots populated.

⚠️ The full pycaffe test suite requires a compiled caffe with GPU support and could not be run locally; please rely on CI for end-to-end validation.

Why existing tests missed it

python/caffe/test/ has no DetectNet clustering coverage, and every shipped DetectNet example happens to produce fewer than 50 boxes per image.