Research map / Detection and segmentation / Tracking
OneTrack: Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers
Qitai Wang, Jiawei He, Yuntao Chen, Zhaoxiang Zhang
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 gradient coordination via four-group query split to joint training (ablation) | OneTrack · nuScenes val (half-res) | AMOTA: 0.425 -> 0.441 | Improved |
| add dynamic attention mask between conflicted query groups (ablation) | OneTrack · nuScenes val (half-res) | AMOTA: 0.441 -> 0.469 | Improved |
| add high-cost assignment suppression (modified tracking loss) (ablation) | OneTrack · nuScenes val (half-res) | AMOTA: 0.469 -> 0.492 AMOTA; NDS 0.575 -> 0.566 | Mixed |
| train tracker only versus detector only (ablation) | End-to-end tracker · nuScenes val (half-res) | NDS: 0.570 -> 0.455 NDS (tracking only) | Hurt |
| initialize tracker from pre-trained detector (ablation) | End-to-end tracker · nuScenes val (half-res) | AMOTA: 0.305 -> 0.375 | Improved |
| supervise detector with extra tracking classification loss on ground truth assignments (ablation) | 3D detector · nuScenes val (pilot) | NDS: 0.570 -> 0.512 NDS; mAP 0.479 -> 0.432 | Hurt |
| use separated classification heads for detection and tracking (ablation) | OneTrack · nuScenes val (half-res) | AMOTA: 49.0 -> 49.2 | Improved |
| increase track query max age Aage (ablation) | OneTrack · nuScenes val (half-res) | IDS | Hurt |
| increase track query init threshold Tconf (ablation) | OneTrack · nuScenes val (half-res) | AMOTA/IDS | Mixed |
| OneTrack single-stage training versus DQTrack/PF-Track multi-stage | OneTrack · nuScenes 800x320 | training time: 176h -> 87h (vs DQTrack) | Improved |
| use OneTrack-F instead of PF-Track-F on identity switches | OneTrack · nuScenes val | IDS: 181 -> 389 | Hurt |
Later papers that compared with this method
2 later accepted papers report OneTrack as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- SynCL: A Synergistic Training Strategy with Instance-Aware Contrastive Learning for End-to-End Multi-Camera 3D Tracking (NeurIPS 2025 · better on 2 of 2 shared results)
- Exploiting Multimodal Spatial-temporal Patterns for Video Object Tracking (AAAI 2025 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- OneTracker: Unifying Visual Object Tracking with Foundation Models and Efficient Tuning (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)
- SUTrack: Towards Simple and Unified Single Object Tracking (AAAI 2025 · similarity 0.88)
- SynCL: A Synergistic Training Strategy with Instance-Aware Contrastive Learning for End-to-End Multi-Camera 3D Tracking (NeurIPS 2025 · similarity 0.87)
- VastTrack: Vast Category Visual Object Tracking (NeurIPS 2024 · similarity 0.87)
- Towards Category Unification of 3D Single Object Tracking on Point Clouds (ICLR 2024 · similarity 0.86)
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.
