Research map / Detection and segmentation / Tracking
Focusing on Tracks for Online Multi-Object Tracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TrackTrack, an online multi-object tracker with a track-perspective iterative min-distance association (TPA) that uses all detections including NMS-discarded ones, and a track-aware initialization (TAI) that suppresses redundant new tracks; it reports state-of-the-art HOTA on MOT17, MOT20, and DanceTrack.
Problem
Data association and track initialization in online tracking-by-detection MOT: global Hungarian matching and multi-stage cascades underuse low-confidence and NMS-discarded detections, and naive initialization creates spurious tracks in occluded scenes.
Approach
Instead of global cost minimization, each track is matched to its minimum-cost detection in a joint single-stage pass over all detections (high, low, and NMS-deleted high-confidence), iteratively with a tightening threshold; new tracks are initialized only from detections not overlapping active tracks or more confident detections.
Key results
- TrackTrack reaches 67.1 HOTA on MOT17 test, the highest among online trackers in the comparison table.
- On DanceTrack test, TrackTrack achieves 66.5 HOTA and 52.9 AssA, above Hybrid-SORT-ReID (65.7 HOTA).
- On MOT17 validation, TPA plus TAI raises HOTA from 67.3 to 69.1 over the Hungarian baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Track-Perspective-Based Association (TPA) (ablation) | Baseline tracker · MOT17-val | HOTA: 67.3 -> 68.5 | Helped |
| add Track-Aware Initialization (TAI) (ablation) | Baseline tracker · MOT17-val | HOTA: 67.3 -> 67.4 | Helped |
| replace Hungarian assignment with the proposed assignment method (ablation) | TrackTrack · MOT17-val | HOTA: 68.1 -> 69.1 | Helped |
| replace multi-stage association with joint association (ablation) | TrackTrack · MOT17-val | HOTA: 68.3 -> 69.1 | Helped |
| include deleted high-confidence detection results Ddel (ablation) | TrackTrack · MOT17-val | HOTA: 68.6 -> 69.1 | Helped |
Surprising findings
- The Hungarian-based baseline used in the ablation is slower than the proposed joint association (161.52 vs 155.44 FPS on MOT17), despite the proposed method being plain Python.
Limitations
- The assignment is implemented in straightforward Python and has room for optimization, as the authors note.
- Several hyperparameters (penalties, thresholds, reduction term) are tuned on the validation sets, and the TAI NMS threshold trades off certainty against crowd density.
Experimental setup
Datasets: MOT17 MOT20 DanceTrack CrowdHuman Widerperson COCO
Benchmarks: MOT17 MOT20 DanceTrack
Metrics: HOTA IDF1 MOTA AssA FPS
Baselines: Hybrid-SORT-ReID Deep OC-SORT DeconfuseTrack SparseTrack UCMCTrack+ AdapTrack PIA ImprAsso CMTrack SUSHI CoNo-Link StrongSORT++
Builds on: ByteTrack Hybrid-SORT OC-SORT Kalman filter FastReID
Closest papers by meaning
- Hybrid-SORT: Weak Cues Matter for Online Multi-Object Tracking (AAAI 2024 · similarity 0.89)
- More Than Meets the Eye: Enhancing Multi-Object Tracking Even with Prolonged Occlusions (ICML 2025 · similarity 0.89)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.88)
- Towards Generalizable Multi-Object Tracking (CVPR 2024 · similarity 0.88)
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 2024 · similarity 0.88)
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.88)
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.