Multi-Object Tracking — BoostTrack++ Reproduction

Reproduction study of BoostTrack++: Using Tracklet Information to Detect More Objects in Multiple Object Tracking
Stanojević & Todorović, arXiv:2408.13003, 2024
Original repository: vukasin-stanojevic/BoostTrack

---

Overview

BoostTrack++ extends the BoostTrack tracker with three plug-and-play improvements to the detection confidence boosting step:

• Soft BIoU — tracklet-confidence-aware buffered IoU that scales bounding box expansion proportionally to prediction uncertainty, replacing fixed-threshold IoU
• Soft detection confidence boost — combines original detection confidence with tracklet similarity instead of treating all low-confidence detections equally
• Varying similarity threshold — tracklet-specific thresholds that decay linearly as frames since last update increase, accounting for degrading Kalman filter prediction quality

Each addition is independent and can be applied to any tracking-by-detection (TBD) algorithm. Combined with the BoostTrack+ baseline, the method achieves state-of-the-art HOTA and IDF1 on MOT20.

---

Reproduction Results (MOT17)

Evaluated on the MOT17 benchmark against the BoostTrack baseline.

Method	HOTA	MOTA	IDF1	ID Switches
BoostTrack (baseline)	65.4	80.0	80.6	1086
BoostTrack++ (reproduced)	66.6	80.7	82.2	1062
Gain	+1.2	+0.7	+1.6	-24 (~20% fewer)

Results match the reported values in the original paper (Table 4).

---

Method Summary

Problem

Standard detection confidence boosting (DLO boost in BoostTrack) uses only IoU between detections and tracklets to decide which low-confidence detections to keep. This has three weaknesses:

1. IoU alone is unreliable — ghost tracks and irregular motion cause false positives and misses
2. Low and high confidence detections are treated equally — a detection at 0.05 requires the same IoU as one at 0.35 to get boosted
3. A fixed similarity threshold ignores that unmatched tracklets have degrading prediction quality over time

Solutions

Soft BIoU (Section 4.1): Scales each tracklet's bounding box by (1 - confidence) / 2, so uncertain tracklets get larger search regions. Reduces to standard IoU when tracklet confidence = 1.

Soft confidence boost (Section 4.3): New confidence score computed as:

c_new = max(c_orig, α * c_orig + (1 - α) * S(detection, tracklet)^q)

Parameters: q = 1.5, α = 0.65 (grid-searched on MOT17 validation).

Varying threshold (Section 4.4): Threshold decays linearly from β_high = 0.95 to β_low = 0.80 over 20 frames since last tracklet update (γ = 0.0075).

Combined similarity measure for boosting averages SBIoU, Mahalanobis distance, and shape similarity — requiring all three to agree before boosting a detection's confidence.

---

Setup

Clone the original BoostTrack++ repository and follow their setup instructions:

git clone https://github.com/vukasin-stanojevic/BoostTrack
cd BoostTrack
pip install -r requirements.txt

Download MOT17 dataset from MOT Challenge and place under data/MOT17/.

Download YOLOX-X detector weights (used in the paper, from ByteTrack):

# weights go in: detector_weights/

Run BoostTrack++ on MOT17:

python main.py --dataset mot17 --use_sb --use_vt

Evaluate with TrackEval:

python scripts/run_mot_challenge.py \
  --BENCHMARK MOT17 \
  --SPLIT_TO_EVAL test \
  --TRACKERS_TO_EVAL BoostTrackPP

---

Datasets

Dataset	Sequences	Frames	Notes
MOT17	7 train / 7 test	5316 / 5919	Static and moving camera, pedestrian tracking
MOT20	4 train / 4 test	8931 / 4479	Crowded scenes, changing lighting

---

Key Metrics

• HOTA (Higher Order Tracking Accuracy) — combined detection, association, and localization
• MOTA (Multi-Object Tracking Accuracy) — penalizes false positives and false negatives
• IDF1 — association performance measure
• IDSW (Identity Switches) — number of times a tracked identity changes ID

---

Reference

@article{stanojevic2024boosttrackpp,
  title={BoostTrack++: Using Tracklet Information to Detect More Objects in Multiple Object Tracking},
  author={Stanojević, Vukašin and Todorović, Branimir},
  journal={arXiv preprint arXiv:2408.13003},
  year={2024}
}

---

Stack

PyTorch · YOLOX · Kalman Filter · Hungarian Algorithm · TrackEval · MOT17