Research map / Point clouds / Shape and surface completion
Robust 3D Tracking with Quality-Aware Shape Completion
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 SCVTrack, a voxelized LiDAR 3D single object tracker that maintains a dense synthetic target point cloud via quality-aware shape completion and refines boxes with it.
Problem
3D single object tracking in LiDAR point clouds is hard due to sparse and incomplete target points, and existing dense-feature learning or multi-template aggregation strategies give imprecise shape representations.
Approach
Fuse real target points from historical frames into a synthetic dense target representation via coordinate transform, gating each voxel's fusion with a learned point-cloud quality evaluator so noisy historical predictions do not accumulate; then track with voxelized relation modeling and refine boxes using the synthetic representation.
Key results
- On KITTI, the method reaches a mean success/precision of 65.1/84.5 across four categories, below MTM (70.9/88.4) and MBPT (70.3/87.9) which it does not surpass on mean.
- On NuScenes it achieves mean 52.1 success / 64.7 precision, a 2.9-point success gain over M2T.
- On sparse KITTI scenes it achieves mean 62.0 success / 80.1 precision, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add naive shape completion (ablation) | SCVTrack baseline (BL) · KITTI Car | success: 63.0 -> 65.2 | Helped |
| Replace naive shape completion with quality-aware shape completion (ablation) | SCVTrack baseline with shape completion · KITTI Car | success: 65.2 -> 66.7 | Helped |
| Add quality-aware shape completion (ablation) | SCVTrack baseline (BL) · KITTI Cyclist | success: 72.5 -> 75.1 | Helped |
| Add box refinement to the naive shape completion variant (ablation) | SCVTrack with naive shape completion · KITTI Car | success: 65.2 -> 67.0 | Helped |
| Add box refinement to the quality-aware shape completion variant (ablation) | SCVTrack with quality-aware shape completion · KITTI Car | success: 66.7 -> 68.7 | Helped |
| Use point features instead of voxelized features (ablation) | SCVTrack · KITTI Car | success: 68.7 -> 64.5 | Hurt |
| Use point features instead of voxelized features (ablation) | SCVTrack · KITTI Cyclist | success: 77.4 -> 74.5 | Hurt |
| Use point features instead of voxelized features (ablation) | SCVTrack · KITTI Van | success: 58.6 -> 55.2 | Hurt |
| Use point features instead of voxelized features (ablation) | SCVTrack · KITTI Trailer | success: 59.5 -> 57.5 | Hurt |
| Use SCVTrack instead of M2T | 3D tracking model · NuScenes mean | success: 49.2 -> 52.1 | Helped |
| Use SCVTrack instead of CXT | 3D tracking model pretrained on KITTI · WOD Protocol I mean | success: 42.2 -> 44.8 | Helped |
| Use SCVTrack instead of M2T | 3D tracking model · KITTI mean | success: 70.9 -> 65.1 | Hurt |
Surprising findings
- On the full KITTI benchmark the method's mean success (65.1) is lower than several recent trackers such as MTM and MBPT, despite the paper's claims of favorable performance.
Limitations
- Synthetic target representation is not fully dense in extremely sparse scenes, so gains are limited there.
- Tracking performance trails some recent transformer-based trackers (CXT, MBPT, MTM) on several KITTI categories.
Experimental setup
Datasets: KITTI NuScenes Waymo Open Dataset
Benchmarks: KITTI NuScenes Waymo Open Dataset
Metrics: success AUC precision AUC
Baselines: SC3D P2B LTTR BAT PTT PTTR V2B CMT STNet M2T TAT CXT
Builds on: M2Track V2B TAT PointNet++ BAT
Closest papers by meaning
- 3D Siamese Voxel-to-BEV Tracker for Sparse Point Clouds (NeurIPS 2021 · similarity 0.89)
- Leveraging Shape Completion for 3D Siamese Tracking (CVPR 2019 · similarity 0.89)
- Robust, Real-Time 3D Tracking of Multiple Objects With Similar Appearances (CVPR 2016 · similarity 0.87)
- FSC: Few-point Shape Completion (CVPR 2024 · similarity 0.86)
- Prior-free 3D Object Tracking (CVPR 2025 · similarity 0.85)
- Towards Category Unification of 3D Single Object Tracking on Point Clouds (ICLR 2024 · 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.