Research map / Detection and segmentation / Tracking
3DMOTFormer: Graph Transformer for Online 3D Multi-Object Tracking
Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, Juergen Gall
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 |
|---|---|---|---|
| train with autoregressive sequences of length 6 instead of two-frame pairs (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.6741 -> 0.7121 | Improved |
| increase training length beyond T=6 (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7121 -> 0.7124 | No clear effect |
| use greedy matching instead of Hungarian matching (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7006 -> 0.7121 | Improved |
| use autoregressive tracks instead of GT identity guided association in training (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7073 -> 0.7121 | Improved |
| pass hidden state of previous frame as initial track feature (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7013 -> 0.7121 | Improved |
| add edge features to graph transformer (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.6469 -> 0.7121 | Improved |
| add predicted center distance in edge features (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7055 -> 0.7121 | Improved |
| add frame difference to edge features (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7113 -> 0.7121 | No clear effect |
| use edge-augmented graph transformer instead of node-feature affinity (cosine) (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.6346 -> 0.7121 | Improved |
| use sparse dynamic graph instead of fully connected graph (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7015 -> 0.7121 | Improved |
| halve the graph distance threshold (ablation) | 3DMOTFormer · nuScenes val | AMOTA: 0.7121 -> 0.6952 | Hurt |
| train on one detector and test on another | 3DMOTFormer · nuScenes val (target CenterPoint) | AMOTA: 0.712 -> 0.699 (BEVFusion source), 0.697 (MEGVII source) | Mixed |
Later papers that compared with this method
2 later accepted papers report 3DMOTFormer as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Delving into Dynamic Scene Cue-Consistency for Robust 3D Multi-Object Tracking (AAAI 2026 · better on 5 of 5 shared results)
- GRAE-3DMOT: Geometry Relation-Aware Encoder for Online 3D Multi-Object Tracking (CVPR 2025 · better on 4 of 4 shared results)
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
- M3SOT: Multi-Frame, Multi-Field, Multi-Space 3D Single Object Tracking (AAAI 2024 · similarity 0.90)
- S2-Track: A Simple yet Strong Approach for End-to-End 3D Multi-Object Tracking (ICML 2025 · similarity 0.89)
- GRAE-3DMOT: Geometry Relation-Aware Encoder for Online 3D Multi-Object Tracking (CVPR 2025 · similarity 0.89)
- TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking (NeurIPS 2026 · similarity 0.89)
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.89)
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 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.
