perf(ocsort): reuse the box returned by predict

[JESUSROYETH]

OC-SORT currently throws away the box returned by predict(), then decodes the same unchanged state through get_state_bbox() before association. This happens because the shared predict helper discards return values at base.py:520-525, and OCSORT decodes again at tracker.py:266-275.

So the returned predictions end up optional in the shared helper. OCSORT asks for them and keeps an id(tracklet) map, so pruning tracks before association cannot move the boxes to a different track. Duplicate timestamps still use get_state_bbox() because prediction is skipped there. The other tracker families use the default path and keep the old loop without building the map.

Validation

I compared develop against itself first, then against this patch. Two 500-frame dynamic streams cover XCYCSR and XYXY, 10,547 detections per representation, duplicate timestamps every 37 frames, and a separate case where one track is pruned between prediction and association. Output boxes were paired by Hungarian IoU, the output IDs and complete per-frame state hashes are exact.

In both cases, the same check passes over 36,928 public MOT17 detections and 45,442 DanceTrack ground-truth boxes. It removes 50,563 redundant decodes on MOT17 and 48,874 on the DanceTrack sample.

pytest -m 'not integration' passes with 1,564 tests, the four evaluator integrations pass, and pre-commit run --all-files is clean.

Performance

CPU-only, Python 3.12.3 and NumPy 1.26.4. Inputs are loaded before timing, run order alternates and garbage collection runs before each sample.

stream	develop median [range]	this patch median [range]	delta	paired wins
MOT17, 7 sequences / 2,664 frames	1.615897 s [1.593757, 1.673104]	1.555865 s [1.532014, 1.600218]	-3.72%	10/10
DanceTrack, 5 stratified sequences / 3,676 frames	1.858677 s [1.854952, 1.869702]	1.799468 s [1.793562, 1.818832]	-3.19%	10/10

I also checked the five callers that do not need returned predictions. A first version built the map unconditionally and made them 0.31-0.57% slower, so I dropped it. With opt-in collection their final MOT17 median deltas are between -0.17% and +0.09%, the ranges overlap, and paired wins are mixed at 4-7/10.

Dynamic prediction arrived in #446. #527 optimized the adjacent OCSORT association path, while #528 cached predicted boxes for BoT-SORT and CBIoU. OCSORT only has one primary predicted-box association stage, so its already returned value stayed unused.

[Copilot]

Pull request overview

Optimizes OC-SORT association by reusing bounding boxes returned during prediction instead of decoding unchanged states again.

Changes:

• Adds opt-in prediction collection to the shared helper.
• Reuses predictions by tracklet identity, with duplicate-timestamp fallback.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/trackers/core/base.py	Optionally returns predictions keyed by tracklet identity.
src/trackers/core/ocsort/tracker.py	Reuses cached predictions during association.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.