perf(kf): construct small diagonal covariances directly

[JESUSROYETH]

The scale-aware noise paths in BoTSORTTracklet and McByteTracklet rebuild small diagonal Q/R matrices on predict/update, and P when a track starts. The diagonal values are already known, but each of the twelve sites goes through the general-purpose np.diag dispatch anyway.

So this adds the same small helper to both sibling modules: allocate the zero matrix, then assign its flat diagonal stride directly. It keeps the existing values and the float64 output, it just skips the generic path.

Validation

I first compared the baseline against itself, then against this patch, over 6,000 boxes. That covers BoT-SORT and McByte tracklets, XCYCWH/XYXY/XCYCSR representations, restored float32 and float64 states, initial Q/R/P, predict and update. The complete bit-level digest is identical.

I also ran the three tracker families over 36,928 real MOT17 public detections, plus BoT-SORT over 45,442 real DanceTrack ground-truth boxes from five sequences picked to cover a range of sequence sizes. Output boxes were paired by Hungarian IoU, not by index. Matched boxes, tracker IDs and the complete per-frame state digests are exact.

The focused tracker suite passes with 150 tests. pytest -m 'not integration' passes with 1,564 tests, all 4 evaluator integrations pass, and pre-commit run --all-files is clean.

Performance

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

Isolated construction, 50,000 calls per sample:

matrix	np.diag median [range]	direct median [range]	delta	paired wins
4×4	0.762 µs [0.752, 0.772]	0.515 µs [0.509, 0.521]	-32.33%	15/15
7×7	0.835 µs [0.818, 0.853]	0.571 µs [0.563, 0.594]	-31.53%	15/15
8×8	0.857 µs [0.848, 0.890]	0.602 µs [0.593, 0.620]	-29.67%	15/15

Whole tracker, 10 paired runs:

stream / tracker	develop median [range]	this patch median [range]	delta	paired wins
MOT17 / BoT-SORT	516.62 µs/frame [509.58, 523.09]	502.53 µs/frame [497.52, 509.46]	-2.73%	10/10
MOT17 / CBIoU	527.84 µs/frame [506.98, 532.39]	516.65 µs/frame [497.57, 518.49]	-2.12%	10/10
MOT17 / McByte	593.26 µs/frame [581.69, 596.65]	580.57 µs/frame [569.95, 585.94]	-2.14%	10/10
DanceTrack / BoT-SORT	434.60 µs/frame [428.52, 442.10]	425.34 µs/frame [420.33, 436.75]	-2.13%	9/10

The duplicated noise code came through BoT-SORT #386 and McByte #388. #567 made Q and R refresh only where each is actually consumed, then #572 removed the larger bounding-box decode overhead. That left these small matrix builders as a measurable part of what's left.

Open feature PRs #511 and #436 also touch BoT-SORT's tracklet, so this patch doesn't apply cleanly to their current heads. Neither one changes McByte's sibling or proposes this diagonal construction, but if either lands first the BoT-SORT half will need conflict resolution.

[Copilot]

Pull request overview

Optimizes scale-aware Kalman covariance construction while preserving float64 values and tracker behavior.

Changes:

• Adds direct diagonal-matrix construction.
• Applies it to Q, R, and initial P in BoT-SORT and McByte.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/trackers/core/botsort/tracklet.py	Optimizes BoT-SORT covariance construction.
src/trackers/core/mcbyte/tracklet.py	Applies the same optimization to McByte.

---

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