Fasttrack Eval

fasttrack-eval

FastTracker: Real-Time and Accurate Visual Tracking — Hashempoor et al. (2025) (arXiv:2508.14370, 2025)

What this evaluates

Evaluates multi-object tracking performance in highly crowded, complex urban traffic environments. It probes a tracker's ability to maintain identity consistency under severe occlusions, varying lighting conditions, and across diverse object classes using motion and structural cues rather than appearance models.

Datasets

• FastTrack — total 800000; splits: test (-1); repo https://github.com/Hamidreza-Hashempoor/FastTracker

Metrics

• HOTA (primary) — range: [0, 1]
  • Higher Order Tracking Accuracy (HOTA) jointly evaluates detection and association performance by computing the harmonic mean of Detection Accuracy (DA) and Association Accuracy (AA) across multiple IoU thresholds.

Input / output format

Input: Sequential video frames containing densely populated traffic scenes with multiple object classes, provided as ground-truth annotated sequences for evaluation.

Output: Per-frame predicted bounding boxes, track IDs, and class labels for all tracked objects.

Scoring recipe

def compute_hota(predictions, ground_truth):
    # 1. Match predicted boxes to GT boxes using IoU threshold sweep (0.5 to 0.95)
    # 2. Compute Detection Accuracy (DA) = TP / (TP + FN + FP)
    # 3. Compute Association Accuracy (AA) = TP / (TP + FN)
    # 4. HOTA = sqrt(DA * AA) averaged over all IoU thresholds
    # 5. Return mean HOTA score across frames/videos
    return standard_mot_hota_eval(predictions, ground_truth)

Common pitfalls

• Trackers relying solely on CNN-based appearance features often fail under severe occlusion and rapid lighting changes (daylight to night/shadows).
• High object density (~43.5 objects/frame) causes frequent ID switches and missed detections if motion/structural priors are ignored.
• Multi-class evaluation (9 classes) requires correct class association, not just pedestrian tracking, which many MOT benchmarks overlook.

Evidence (verbatim from paper)

To comprehensively evaluate multi-object tracking in complex traffic scenes, we introduce the FastTrack benchmark—a diverse and challenging dataset that surpasses existing benchmarks such as UrbanTracker and CityFlow in several key dimensions. FastTrack contains 800K annotated detections across 12 videos, each densely populated with an average of 43.5 objects per frame—more than 5× that of UrbanTracker and over 5× CityFlow—making it particularly suitable for evaluating trackers under extreme crowding and interaction. achieving strong HOTA scores (66.4 on MOT17, 65.7 on MOT20) while introducing a new benchmark with diverse vehicle and pedestrian classes in complex urban CCTV scenes.

Citation

@misc{hashempoor2025fasttracker,
  title={FastTracker: Real-Time and Accurate Visual Tracking},
  author={Hashempoor et al. (2025)},
  year={2025},
  note={arXiv:2508.14370}
}

• arXiv: 2508.14370