Research map / Detection and segmentation / Tracking
ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association
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 ADA-Track, an end-to-end multi-camera 3D MOT framework that alternates detection (query-to-image cross-attention) and data association (edge-augmented query-to-query cross-attention) across stacked DETR3D/PETR decoder layers.
Problem
Query-based end-to-end multi-camera 3D multi-object tracking suffers from entangled detection/tracking queries in tracking-by-attention and from decoupled, sequentially processed detection and association in tracking-by-detection, so the two tasks are not optimized jointly.
Approach
Decouple track and detection queries but integrate a learned edge-augmented cross-attention association module (with appearance node features and geometric edge features) into every transformer decoder layer of a DETR-style 3D detector, so detection and association refine queries alternately and iteratively.
Key results
- On nuScenes val with DETR3D, ADA-Track reaches 0.378 AMOTA, 2.8 points above the TBD-Baseline (0.350).
- With PETR on nuScenes val, ADA-Track achieves 0.479 AMOTA, matching PF-Track (0.479) and 2.6 points above PF-Track without track extension (0.453).
- On nuScenes test, ADA-Track reaches 0.456 AMOTA, outperforming STAR-Track (0.439) and PF-Track (0.434).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the TBA paradigm with alternating detection and association | DETR3D tracking · nuScenes validation | AMOTA: 0.321 -> 0.378 | Helped |
| replace the TBD paradigm with alternating detection and association | DETR3D tracking · nuScenes validation | AMOTA: 0.350 -> 0.378 | Helped |
| replace the TBD paradigm with alternating detection and association | PETR tracking · nuScenes validation | AMOTA: 0.452 -> 0.479 | Helped |
| use ADA-Track instead of PF-Track | nuScenes tracking · nuScenes test | AMOTA: 0.434 -> 0.456 | Helped |
| increase training sample length from T=2 to T=3 (ablation) | ADA-Track · nuScenes validation | AMOTA: 0.346 -> 0.378 | Helped |
| increase training sample length from T=3 to T=4 (ablation) | ADA-Track · nuScenes validation | AMOTA: 0.378 -> 0.379 | No clear effect |
| use association outputs from later decoder layers with final-layer boxes (ablation) | ADA-Track · nuScenes validation | AMOTA: 0.342 -> 0.369 -> 0.372 -> 0.373 -> 0.374 | Helped |
| use boxes from later decoder layers with final-layer association outputs (ablation) | ADA-Track · nuScenes validation | AMOTA: 0.279 -> 0.364 -> 0.368 -> 0.373 -> 0.373 -> 0.378 | Helped |
| set feature update weight wt to 0 instead of a nonzero value (ablation) | ADA-Track · nuScenes validation | AMOTA: 0.367 -> 0.378; 0.371 -> 0.378; 0.363 -> 0.378; 0.365 -> 0.378 | Helped |
| use the track-predicted box instead of the detection-predicted box for associated pairs (ablation) | ADA-Track · nuScenes validation | AMOTA: 0.378 -> 0.365 | Hurt |
| reverse edge-augmented cross-attention direction to track queries as queries and detection queries as keys (ablation) | ADA-Track · nuScenes validation | AMOTA: 0.365 -> 0.343 | Hurt |
Surprising findings
- Substituting track query features with detection query features (wt=0) performed best, rather than merging both query types.
- Using track queries as queries and detection queries as keys in the association attention reduced AMOTA by 2.2 points.
Limitations
- Training T=4 frames adds considerable computational and memory overhead, so training sample length is limited to 3 frames.
- Results rely on a frozen image backbone and FPN for DETR3D experiments, and only nuScenes is evaluated.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes validation nuScenes test
Metrics: AMOTA AMOTP recall MOTA IDS FP FN
Baselines: MUTR3D DQTrack STAR-Track PF-Track TBA-Baseline TBD-Baseline CC-3DT QD-3DT DEFT Time3D
Builds on: DETR3D PETR 3DMOTFormer MUTR3D Edge-Augmented Graph Transformer
Closest papers by meaning
- S2-Track: A Simple yet Strong Approach for End-to-End 3D Multi-Object Tracking (ICML 2025 · similarity 0.89)
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 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 0.88)
- GRAE-3DMOT: Geometry Relation-Aware Encoder for Online 3D Multi-Object Tracking (CVPR 2025 · similarity 0.88)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · 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.