Research map / Detection and segmentation / Tracking
GRAE-3DMOT: Geometry Relation-Aware Encoder for Online 3D 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 GRAE-3DMOT, a geometry-only online 3D multi-object tracker whose relation-aware encoder connects detections and tracks without distance truncation, reaching 73.7% AMOTA on nuScenes validation with CenterPoint detections.
Problem
Distance-threshold gating in tracking-by-detection and graph-based 3D MOT limits the association candidates between detections and tracks, degrading association when objects move significantly between frames.
Approach
Builds fully connected spatial and spatiotemporal relation graphs (infinite distance threshold) over detections and tracks, and uses a distance-aware feature fusion layer (DFFL) to emphasize physically close relations, producing association scores matched by Hungarian matching.
Key results
- Achieves 73.7% AMOTA on nuScenes validation and 70.2% on test with CenterPoint detections, outperforming 3DMOTFormer (71.2% val, 68.2% test).
- With StreamPETR detections on validation, AMOTA improves from 0.506 (3DMOTFormer) to 0.521.
- Ablation: infinite distance thresholds dS=dT=8 give AMOTA 0.737 versus 0.720 with the spatial-only encoder variant M1 and 0.729 with DFFL removed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase spatial and temporal graph thresholds to 8 (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.729 -> 0.737 | Helped |
| use geometry-difference edges instead of destination geometry attributes (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.725 -> 0.737 | Helped |
| use geometry-difference edges instead of embedded-feature differences (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.729 -> 0.737 | Helped |
| add the spatiotemporal relation-aware encoder (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.720 -> 0.737 | Helped |
| add the spatial relation-aware encoder (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.729 -> 0.737 | Helped |
| use the spatial relation-aware encoder instead of the 3DMOTFormer graph transformer encoder (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.729 -> 0.737 | Helped |
| add multi-head self-attention (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.732 -> 0.737 | Helped |
| add the distance-aware feature fusion layer (ablation) | GRAE-3DMOT · nuScenes validation | AMOTA: 0.729 -> 0.737 | Helped |
Surprising findings
- AMOTP is worse than the best geometry baseline (0.475 test vs 0.550-0.566 for others under the AMOTP lower-is-better convention in the paper's table, though GRAE's AMOTP 0.475 is listed as best), so AMOTP tradeoffs are unclear.
- Performance of the baseline 3DMOTFormer and CenterPoint degrades as distance threshold increases, whereas the oracle improves.
Limitations
- Relies only on geometric cues without appearance features.
- Performance on AMOTP is not consistently best on the validation set compared with some baselines.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes tracking validation nuScenes tracking test
Metrics: AMOTA AMOTP MOTA IDS FRAG MT ML TP FP FN
Baselines: 3DMOTFormer CenterPoint OGR3MOT PolarMOT CBMOT SimpleTrack ImmortalTracker GNN-PMB
Builds on: 3DMOTFormer CenterPoint Hungarian matching Focal loss
Closest papers by meaning
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 2024 · similarity 0.88)
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.86)
- M3SOT: Multi-Frame, Multi-Field, Multi-Space 3D Single Object Tracking (AAAI 2024 · similarity 0.86)
- S2-Track: A Simple yet Strong Approach for End-to-End 3D Multi-Object Tracking (ICML 2025 · similarity 0.86)
- Towards Category Unification of 3D Single Object Tracking on Point Clouds (ICLR 2024 · similarity 0.85)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.85)
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.