roboflow--rf-detr

Files

T

History

Jesús Royeth 931245dbf5 perf(matcher): compute detection-only cost matrix padded to max(T_i), not sum(T_i) (#1297 )

- Improved Hungarian matching performance by using a compact detection cost matrix that scales with per-image target counts while preserving assignment behavior and falling back to the full path for unsafe or incompatible inputs
- Improved matcher validation for target dtype/device mismatches, out-of-range labels, non-finite relevant costs, extreme coordinates, grouped-query constraints, and float64 inputs while preserving established full-path behavior
- Refactored focal classification cost calculation so compact and full Hungarian matching paths share the same formula
- Fixed concurrent COCO dataset downloads by serializing access per asset, using process-unique staging files, retrying truncated archives, cleaning up partial downloads, and avoiding redundant downloads after another process completes extraction

---------

Co-authored-by: Jesús Royeth <JESUSROYETH@users.noreply.github.com>
Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

2026-08-07 20:44:14 +02:00

conftest.py

perf(matcher): compute detection-only cost matrix padded to max(T_i), not sum(T_i) (#1297 )

2026-08-07 20:44:14 +02:00

test_develop_downloads.py

perf(matcher): compute detection-only cost matrix padded to max(T_i), not sum(T_i) (#1297 )

2026-08-07 20:44:14 +02:00

test_develop_urls.py

feat(datasets): add COCO keypoint loading and aug-warning guard

2026-06-12 15:12:07 +02:00

test_inference_coco.py

Add doctest coverage to benchmark and dataset test helpers (#1296 )