Research map / Label-efficient and robust learning / Recognition and long-tail
SPAMming Labels: Efficient Annotations for the Trackers of Tomorrow
Orcun Cetintas, Tim Meinhardt, Guillem Brasó, Laura Leal-Taixé
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add low-confidence detection boxes without GNN_node filtering (ablation) | SPAM graph tracker · MOT17 val half | HOTA: 59.9 -> 58.5 | Hurt |
| add GNN_node detection-filtering level on top of low-confidence boxes (ablation) | SPAM graph tracker · MOT17 val half | HOTA: 58.5 -> 60.4 | Improved |
| retrain model on its own pseudo-labels after synthetic pretraining (ablation) | SPAM labeler · MOT17 val (ByteTrack trained on labels) | HOTA: 49.6 -> 50.5 | Improved |
| add uncertainty-based active learning with 3.3% manual effort (ablation) | SPAM labels · MOT17 val (ByteTrack trained on labels) | HOTA: 50.5 -> 52.5 | Improved |
| pseudo-label retraining on real data (zero human effort) (ablation) | SPAM · MOT17, MOT20, DanceTrack | HOTA/MOTA/IDF1: +4-6 HOTA, +4-6 MOTA, +4-7 IDF1 | Improved |
| box-level graph-based active learning instead of image-level or random labeling (ablation) | SPAM active learning · MOT17 val | HOTA gain over no-AL baseline: 1.2 points from Oracle | Improved |
| use Coreset (diversity) selection instead of uncertainty-based selection (ablation) | SPAM active learning · MOT17 val | HOTA gain | Hurt |
| train detector on real instead of synthetic data (ablation) | tracking components · MOT17/MOT20/DanceTrack | HOTA: 9.9 HOTA | Improved |
| train reID on real instead of synthetic data (ablation) | tracking components · MOT17/MOT20/DanceTrack | HOTA | No clear effect |
| train association on real instead of synthetic data (ablation) | tracking components · MOT17/MOT20/DanceTrack | HOTA: +2.1 HOTA | Improved |
| SPAM labels at 3.3% effort vs ground truth on harder datasets | ByteTrack and GHOST · MOT20 and DanceTrack test | HOTA: ByteTrack DanceTrack 39.5 vs 41.0 (GT) | Mixed |
Closest papers by meaning
- PathTrack: Fast Trajectory Annotation With Path Supervision (ICCV 2017 · similarity 0.89)
- TAO: A Large-Scale Benchmark for Tracking Any Object (ECCV 2020 · similarity 0.87)
- Track-On2: Enhancing Online Point Tracking With Memory (TPAMI 2026 · similarity 0.86)
- Track-On: Transformer-based Online Point Tracking with Memory (ICLR 2025 · similarity 0.86)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.84)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
