Research map / 3D vision / Human pose estimation
Towards Category Unification of 3D Single Object Tracking on Point Clouds
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 category-unified 3D single object tracking models (SiamCUT and MoCUT) that track all object categories with one network, using the AdaFormer deformable group vector-attention backbone, size-scaled search regions, and size-normalized offset targets with shape-aware labels.
Problem
Existing 3D single object tracking methods train and evaluate a separate model per object category on point clouds, causing redundant parameters and limited generalization; training one model on cross-category data is unstable and suboptimal.
Approach
Encode category-specific size and shape variation adaptively via a deformable group vector-attention network (AdaFormer), build search regions by scaling the previous box by its own dimensions so foreground-background ratios are consistent, and normalize offset targets and positive labels by the template's width, height, and length so one model learns a unified objective.
Key results
- MoCUT reaches 65.8% Success and 85.0% Precision mean on KITTI with a single model, versus 62.9/83.4 for category-specific M2Track.
- SiamCUT improves on category-specific P2B from 42.4/60.0 to 54.0/74.6 mean on KITTI, about 11.6 Success points.
- On NuScenes, MoCUT achieves 51.19/64.63 mean versus 49.32/62.73 for M2Track; SiamCUT achieves 40.41/48.54 versus P2B's 36.48/45.08.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use category-unified SiamCUT instead of category-specific P2B | P2B · KITTI | Success: 42.4 -> 54.0 | Helped |
| Use category-unified MoCUT instead of category-specific M$^2$Track | M$^2$Track · KITTI | Success: 62.9 -> 65.8 | Helped |
| Use category-unified SiamCUT instead of category-specific P2B | P2B · NuScenes | Success: 36.48 -> 40.41 | Helped |
| Use category-unified MoCUT instead of category-specific M$^2$Track | M$^2$Track · NuScenes | Success: 49.32 -> 51.19 | Helped |
| Add AdaFormer (ablation) | SiamCUT · KITTI | Success: 45.9 -> 54.0 | Helped |
| Add unified model inputs (ablation) | SiamCUT · KITTI | Success: 49.1 -> 54.0 | Helped |
| Add unified learning objective (ablation) | SiamCUT · KITTI | Success: 51.8 -> 54.0 | Helped |
| Add AdaFormer (ablation) | MoCUT · KITTI | Success: 63.5 -> 65.8 | Helped |
| Add unified model inputs (ablation) | MoCUT · KITTI | Success: 65.5 -> 65.8 | Helped |
| Add unified learning objective (ablation) | MoCUT · KITTI | Success: 65.1 -> 65.8 | Helped |
Surprising findings
- Naively training category-specific trackers on cross-category data yields unstable performance after convergence, and the unified components reduce this jitter.
- The Siamese-paradigm gains are larger than the motion-centric gains, which the authors attribute to the Siamese paradigm's reliance on shape and size embedding.
Limitations
- Gains for the Siamese paradigm are not uniform: SiamCUT is below category-specific baselines on some subsets in the KITTI category-level comparison.
- The unified model is still trained and tested per dataset (KITTI or NuScenes), so cross-dataset category unification is only partly tested via Waymo evaluation with a KITTI-trained model.
Experimental setup
Datasets: KITTI NuScenes Waymo Open Dataset
Benchmarks: KITTI NuScenes Waymo Open Dataset
Metrics: Success Precision FLOPs parameters FPS
Baselines: P2B M2Track SC3D MLVSNet PTT V2B PTTR TAT STNet DMT GLT-T OSP2B
Builds on: P2B M2Track PointNet++ VoteNet vector attention
Closest papers by meaning
- Generalizable Structure-Aware Keypoint Correspondence for Category-Unified 3D Single Object Tracking (CVPR 2026 · similarity 0.91)
- PointRePar : SpatioTemporal Point Relation Parsing for Robust Category-Unified 3D Tracking (ICLR 2026 · similarity 0.91)
- M3SOT: Multi-Frame, Multi-Field, Multi-Space 3D Single Object Tracking (AAAI 2024 · similarity 0.90)
- Modeling Continuous Motion for 3D Point Cloud Object Tracking (AAAI 2024 · similarity 0.87)
- SUTrack: Towards Simple and Unified Single Object Tracking (AAAI 2025 · similarity 0.86)
- One for All: Multi-Domain Joint Training for Point Cloud Based 3D Object Detection (NeurIPS 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.