VisionTrack

A from-scratch multi-object tracker, used as a controlled study of when the field's standard tricks actually help.

The tracker — an 8-state Kalman filter, an O(n³) Hungarian solver, and ByteTrack two-stage association — is implemented from first principles on NumPy, with no ML framework in the core. On top of it sits a reproducible experiment harness that measures, on real MOT17 with seed variance and paired significance tests, whether appearance and uncertainty-aware association actually improve tracking. Several of the answers are honest negatives — which is the point.

The tracker on a synthetic scene — boxes in, stable per-object IDs out. A real-MOT17 version can be rendered locally on your own copy of the dataset with python scripts/render_mot17_demo.py (the frames aren't redistributed here — MOT17 is under a non-commercial share-alike licence).

Abstract

Most tracking repositories are a thin wrapper over a detector plus a vendored tracker; every reported number is a single run on one configuration. VisionTrack inverts that: the estimation and association math is the deliverable (independently tested against SciPy and trackeval), and it is used to run a falsifiable study. We reproduce a from-scratch ByteTrack baseline on MOT17 public detections (HOTA/IDF1 verified against trackeval to within 1.4e-3), then ablate two common enhancements under seed variance and Wilcoxon significance:

• Appearance (RQ1) — a per-track re-ID cost helps association on MOT17 (ID switches 188 → 170) but the effect is small and not significant with a cheap descriptor.
• Uncertainty (RQ3) — folding calibrated Kalman uncertainty into the cost is null, and calibrating the filter to real motion is actively harmful under detector noise (ID switches 40 → 400+). The filter's apparent under-confidence turns out to be a robustness feature, not a bug.

Synthetic sanity check: seven objects over 120 frames — tracks #3 and #7 cross in the centre and keep their identities.

Research questions

The value is the same whichever way each result falls: a controlled study that shows a trick doesn't help is a contribution, not a failure.

Method

The from-scratch core (core/, NumPy only):

• Kalman filter — 8-state constant-velocity model in DeepSORT's xyah parametrization, height-scaled noise for scale-invariance, Joseph-form covariance update, Mahalanobis gating. Convergence-tested.
• Hungarian assignment — rectangular O(n³) Kuhn–Munkres with dual potentials; validated against scipy.optimize.linear_sum_assignment on 150 random matrices.
• ByteTrack two-stage association + a Tentative → Confirmed → Deleted lifecycle FSM.

The ablation surface (tracking/cost.py): the association cost is factored so each hypothesis is one weighted term with a hard gate deciding feasibility and the terms only ranking feasible pairs (DeepSORT-style fusion):

cost = w_iou·motion  ⊕  w_app·appearance  ⊕  w_unc·uncertainty      (gate: IoU + class + Mahalanobis)

At w_app = w_unc = 0 this is bit-identical to the plain 1 − IoU baseline, so every ablation toggles exactly one variable.

Rigor (eval/, experiments/):

• Metrics: from-scratch CLEAR-MOT + IDF1 + HOTA, cross-checked against trackeval end-to-end on real MOT17 (MOTA/IDF1 exact, HOTA within 1.4e-3).
• Seed variance + paired significance: every configuration is run over seeds; variants are compared paired (same sequences/seeds) with a paired bootstrap + Wilcoxon test and Cohen's d.
• Compute once: detections and appearance embeddings are cached to disk (5.2 MB detections; 6.3 MB colour-histogram or 67 MB deep-re-ID features for all of MOT17-train), so the raw ~5 GB of frames can be deleted and every experiment runs CPU-only in seconds.

Results

Baseline — from-scratch ByteTrack on real MOT17 (val-half, public detections)

Squarely in the public-detection neighbourhood (the famous ~76 MOTA ByteTrack uses a private YOLOX detector), with the expected detector ordering. scripts/xcheck_mot17_trackeval.py confirms our whole pipeline (preprocessing + metrics) matches trackeval on a real sequence.

RQ1 — appearance (MOT17 FRCNN), from-scratch histogram vs deep re-ID

Δ vs the w_app=0 motion-only baseline (HOTA 0.497 / IDF1 0.570 / AssA 0.587 / IDSW 188), same 7-sequence pairing:

Appearance's clearest effect is on ID switches. A deep re-ID embedder — a pretrained OSNet run through appearance/reid_onnx.py behind the same interface — roughly doubles the association gain over the hand-crafted histogram and cuts ID switches to 163 (−13%), earning that weight early (−14 IDSW already at w_app=0.15).

Pooling all three public detectors (DPM/FRCNN/SDP → 21 seq×detector units) for statistical power, the ID-switch and IDF1 reductions become significant (paired Wilcoxon p<0.05); association-quality (AssA/HOTA) stays marginal (p≈0.06). The instructive twist: appearance is completely inert on the weak DPM detector — its poorly-localized boxes yield mis-framed crops, so the re-ID embeddings carry no identity signal. Detection/crop quality gates whether appearance helps at all, the opposite of the "weak detector needs it most" intuition, and it doesn't stratify by crowd density or occlusion. Details →

RQ2 — learned motion residual (helps prediction, hurts tracking)

A from-scratch NumPy MLP (hand-written back-prop + Adam — no framework) predicts a correction to the constant-velocity Kalman mean, trained on GT trajectories.

Open-loop, the residual helps exactly where motion is non-linear — but wired into the tracker it hurts identity on both (DanceTrack HOTA −0.043, +29 IDSW; MOT17 HOTA −0.013, +12 IDSW). Trained on clean GT but fed the tracker's noisy estimates, its correction mis-fires and perturbs the association gate. A better predictor that makes a worse tracker. Details →

RQ3 — calibrated uncertainty (the interesting negative)

Stepping the Kalman filter along real MOT17 GT, the innovation χ² is 0.15 where a calibrated filter would give 4.0 — it is ~25× under-confident, so its 95% gate captures 100% of innovations and never rejects.

Consequences, measured:

• Folding uncertainty into the cost (w_unc) is null (the gate is inert, the signal near-constant).
• Calibrating the filter (kf_noise_scale=0.19) is catastrophic under detector noise: HOTA −0.205, ID switches 40 → 400+ (both p<0.05). A tight gate tuned to clean motion rejects noisy-but-correct detections and tracks fragment.

The loose gate is a robustness feature, not a bug. This unifies with the ablation finding that disabling the gate barely changes clean-data results. Details →

Throughput

Real-time on CPU (single-threaded): 2334 FPS at 4 objects down to 214 FPS at 64 — comfortably ≥30 FPS across the range (benchmarks/bench_tracker.py).

Install & use

pip install visiontrack-mot            # core: NumPy only
pip install 'visiontrack-mot[video]'   # + run on your own video files

from visiontrack import ByteTracker, TrackerConfig
tracker = ByteTracker(TrackerConfig())
for frame_detections in stream:            # list[Detection]
    observations = tracker.update(frame_detections)

Track a real video from the command line (needs a YOLOX ONNX model, see docs/VIDEO.md):

visiontrack track input.mp4 out.mp4 --model models/yolox_nano.onnx

Full public API: docs/API.md · packaging/release: docs/RELEASE.md.

Benchmark a set of trackers into one report (leaderboard + paired significance + ID-switch error taxonomy) — make benchmark, or see the live report at visiontrack.hulage.in/benchmark.

Reproduce

No install? Run the synthetic study in your browser: — clones, installs the harness extra, and reproduces the seed-varied, significance-tested synthetic tables (a couple of minutes, CPU-only). Notebook: notebooks/reproduce.ipynb.

make install                 # editable install with all extras
make test                    # full test suite
make reproduce-synth         # synthetic harness + RQ3 probe — NO dataset needed

Full real-data reproduction (after a one-time cache build — see docs/PHASE0.md):

python data/cache/precompute.py            --data-root ~/MOT17 --detector FRCNN --out data/cache/mot17
python data/cache/precompute_embeddings.py --data-root ~/MOT17 --detector FRCNN --cache-dir data/cache/mot17
make reproduce               # regenerates every table and figure above

Every run is pinned by a config hash; bootstrap resampling and synthetic scenes are seeded. Each phase has a runbook in docs/ (PHASE0.md … PHASE5.md).

Engineering

• Zero ML-framework dependency in the core — just NumPy. Heavy/optional deps (scipy, pandas, matplotlib, pillow, onnxruntime, torch) are isolated in extras ([experiments], [appearance], [onnx]) and lazily imported; nothing in core/ imports them.
• 253 tests: unit, property (Hungarian vs SciPy), convergence (Kalman), metric cross-checks (HOTA/IDF1 vs trackeval), and end-to-end integration with a MOTA floor.
• CI on Python 3.10/3.11/3.12 + ruff.

src/visiontrack/
  core/        geometry · kalman · assignment          ← from-scratch math
  detection/   base · synthetic · onnx_yolo · mot_loader · noise
  appearance/  embedder (colour-hist) · reid_onnx · gallery   (RQ1)
  tracking/    tracker · track (FSM) · cost (ablation surface) · config
  eval/        mot (CLEAR-MOT) · hota (HOTA/IDF1) · stats · calibration (RQ3)
  datasets/    splits (frozen) · cache (detections + embeddings)
experiments/   run_matrix · analyze · appearance_study · uncertainty_study · configs/
data/cache/    precompute · precompute_embeddings
scripts/       xcheck_mot17_trackeval.py

Limitations & honest negatives

• Public-detection, train/val split — reproducible and self-contained, not test-server leaderboard numbers (deliberate).
• RQ1's effect is real but small on MOT17 — a deep re-ID embedder (OSNet-x0.25, MSMT17) roughly doubles the association gain over the from-scratch histogram; pooling all three detectors (21 units) makes the ID-switch/IDF1 reduction significant, but association quality (AssA/HOTA) stays marginal and the benefit is capped by public-detection crop quality. A controlled synthetic probe (dialing inter-object appearance similarity) confirms the benefit grows with object distinctness and is significant on AssA/IDF1 — but never flips to harmful. And the decisive test — real DanceTrack (near-identical dancers, non-linear motion, the predicted "appearance hurts" case) — refutes the hypothesis: deep re-ID still significantly cuts ID switches there (217→202, p<0.05). Across all three probes a gated appearance cost is robustly beneficial-or-neutral. The "appearance hurts on DanceTrack" failure is a property of appearance-vetoing designs, not of uniform appearance itself — our cost only lets appearance rank within the motion gate, never veto a feasible match. Re-ID weights carry a non-commercial dataset licence and are not committed (regenerate from your own download).
• RQ2 (learned motion residual) is answered — an honest negative. A from-scratch NumPy MLP (hand-written back-prop, no framework) lowers open-loop next-centre error on DanceTrack (−12.5%) but not MOT17 (near-linear, ~1px CV error → +21.5% worse); crucially, inside the tracker it hurts on both (train-on-GT / infer-on-noisy-estimates distribution shift + it perturbs the association gate). Trained on GT trajectories; weights gitignored. Details → docs/PHASE4.md
• The cross-dataset RQ1 test is done on DanceTrack, with a scope caveat — detections are oracle-perturbed GT (DanceTrack ships none), which gives clean crops that favour appearance; a poor real detector would weaken it (cf. DPM). The cost is weighted-and-gated, not pure-appearance. So the refutation is scoped: a gated ranking appearance cost does not hurt, even on DanceTrack. SportsMOT (RQ2 maneuver data) is still deferred.

Interactive demo

A self-contained, pre-baked demo that makes the RQ1 result tangible: the same synthetic scene tracked two ways — motion-only vs motion + appearance — with the running ID-switch count side by side. Boxes are coloured by track ID, so a colour flip on a moving object is a switch.

• Try it live: visiontrack.hulage.in/demo — no install, runs in the browser (deployed from viz/webdemo/ via Vercel; project home at visiontrack.hulage.in).
• Build it locally: make demo → open viz/webdemo/index.html in any browser (no server, no dataset, no toolchain — inference is pre-baked from the NumPy tracker on a synthetic scene).
• On the selected scene, appearance cuts ID switches 37 → 29 (−22%) and lifts IDF1 — the study result, watchable frame by frame.

Write-up

A narrative walk-through of the study — why appearance refuses to hurt (even on DanceTrack), how a better motion predictor made a worse tracker, and why the filter's loose gate is a feature: visiontrack.hulage.in/writeup. The honest negatives, explained in plain prose rather than tables.

Study guide

New to the concepts? Two self-paced, interactive learning files (single self-contained HTML, progress checkboxes saved in your browser — open in any browser):

• docs/LEARNING_PATH.html — this project, topic by topic: Kalman → Hungarian → ByteTrack → metrics → the three research questions → where to take it next, with every concept linked to the file it lives in.
• docs/CV_ROADMAP.html — the whole field, in order: a junior → mid → senior → research computer-vision roadmap (foundations → deep learning → detection/segmentation/tracking → transformers/generative/3D → production → doing research), with what-to-build and canonical resources at each stage.

License

MIT