Research map / Detection and segmentation / Tracking
End-to-end 3D Tracking with Decoupled Queries
Yanwei Li, Zhiding Yu, Jonah Philion, Anima Anandkumar, Sanja Fidler, Jiaya Jia, Jose Alvarez
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 |
|---|---|---|---|
| decouple single query into object query and track query (ablation) | DQTrack · nuScenes val | AMOTA: 24.5% -> 28.5% | Improved |
| use single query for detection and tracking instead of tracking-by-detection (ablation) | SinQuery · nuScenes val | AMOTA: 25.7% -> 24.5% | Hurt |
| use DQTrack instead of TBD-Geo with DETR3D | DQTrack · nuScenes val | AMOTA: 32.6% -> 36.7%; IDS 2564 -> 1120 | Improved |
| apply DQTrack across different encoders and decoders | DQTrack · nuScenes val | AMOTA: more than 3.0% AMOTA gain over TBD-Geo | Improved |
| add embedding interaction (ablation) | DQTrack · nuScenes val | AMOTA: 27.8% -> 28.5% | Improved |
| fuse appearance and motion features for association (ablation) | DQTrack · nuScenes val | AMOTA: 26.9% (motion only) -> 28.5% | Improved |
| use transformer decoder embedding instead of raw BEV feature for object embedding (ablation) | DQTrack · nuScenes val | AMOTA: 26.4% -> 28.5% | Improved |
| add EMA temporal update of track query (alpha 0.5) (ablation) | DQTrack · nuScenes val | AMOTA: 27.4% -> 28.5% | Improved |
| add track query augmentation with false positives (ablation) | DQTrack · nuScenes val | AMOTA: 28.1% -> 28.5% | Improved |
| add maximum entropy regularization on unmatched pairs (ablation) | DQTrack · nuScenes val | AMOTA: 27.9% -> 28.5% | Improved |
| add EMA object embedding update (beta 0.3) (ablation) | DQTrack · nuScenes val | AMOTA: 26.9% -> 28.5% | Improved |
| train with 3 frames instead of 2 or 4 (ablation) | DQTrack · nuScenes val | AMOTA: 26.8% (D=2) / 27.0% (D=4) -> 28.5% (D=3) | Improved |
Closest papers by meaning
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 2024 · similarity 0.90)
- OneTrack: Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers (ECCV 2024 · similarity 0.88)
- S2-Track: A Simple yet Strong Approach for End-to-End 3D Multi-Object Tracking (ICML 2025 · similarity 0.87)
- SynCL: A Synergistic Training Strategy with Instance-Aware Contrastive Learning for End-to-End Multi-Camera 3D Tracking (NeurIPS 2025 · similarity 0.85)
- DeconfuseTrack: Dealing with Confusion for Multi-Object Tracking (CVPR 2024 · similarity 0.84)
- GRAE-3DMOT: Geometry Relation-Aware Encoder for Online 3D Multi-Object Tracking (CVPR 2025 · 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.
