Research map / Point clouds / Point cloud classification
Modeling Continuous Motion for 3D Point Cloud 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 StreamTrack, a memory-bank-based streaming Transformer for LiDAR 3D single object tracking that uses hybrid global/local attention and contrastive sequence enhancement with negative tracklets.
Problem
3D single object tracking from LiDAR point clouds typically uses only two successive frames (appearance matching or motion modeling), overlooking long-range continuous motion and suffering from distractors and target switching.
Approach
Treat each tracklet as a continuous stream: only the current frame is encoded per timestep and fused with cached historical point features and box predictions from a memory bank, using a hybrid (global spatial-temporal plus local spatial) attention, query-based decoding, and a contrastive loss over sequences augmented with attached negative tracklets.
Key results
- On KITTI, StreamTrack reaches mean Success 70.8 and Precision 88.1, versus 67.5/85.3 for CXTrack (previous best listed).
- On Waymo, mean Success 55.95 and Precision 71.90 using only 10% of training tracklets, versus 43.13 Success for M2-Track.
- Runs at 40.7 FPS on a single V100 GPU, on par with matching-based methods such as P2B and BAT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use StreamTrack instead of CXTrack | KITTI tracking · KITTI | Success: 67.5 -> 70.8 | Helped |
| use StreamTrack instead of CXTrack | KITTI tracking · KITTI | Precision: 85.3 -> 88.1 | Helped |
| use StreamTrack instead of M²-Track | nuScenes tracking · nuScenes mean | Success: 49.23 -> 55.75 | Helped |
| use StreamTrack instead of M²-Track | nuScenes tracking · nuScenes mean | Precision: 62.73 -> 69.22 | Helped |
| use StreamTrack instead of M²-Track | Waymo tracking · Waymo mean | Success: 43.13 -> 55.95 | Helped |
| use StreamTrack instead of M²-Track | Waymo tracking · Waymo mean | Precision: 63.48 -> 71.90 | Helped |
| add hybrid attention (ablation) | StreamTrack · KITTI mean | Success: 68.9 -> 70.8 | Helped |
| add point supervision (ablation) | StreamTrack · KITTI mean | Success: 65.9 -> 70.8 | Helped |
| use query-based prediction instead of the M²-Track prediction head (ablation) | StreamTrack · KITTI mean | Success: 67.1 -> 68.4 | Helped |
| add sequence enhancement (ablation) | StreamTrack · KITTI mean | Success: 66.5 -> 70.8 | Helped |
| add auxiliary contrastive loss (ablation) | StreamTrack · KITTI mean | Success: 68.0 -> 70.8 | Helped |
| increase input from 2 frames to 3 frames (ablation) | StreamTrack · KITTI | Success: +4.8% | Helped |
Surprising findings
- Motion-centric M2-Track degrades when more historical frames are concatenated, while StreamTrack gains mainly from going from 2 to 3 frames and then plateaus.
Limitations
- Relies on a fixed number of historical frames (n=2 default) and a small search range, with diminishing returns beyond that.
- Performance on Van and Cyclist categories depends on limited training samples; sequence enhancement helps but gains are data-dependent.
Experimental setup
Datasets: KITTI nuScenes Waymo Open Dataset
Benchmarks: KITTI nuScenes Waymo Open Dataset
Metrics: Success Precision FPS
Baselines: SC3D P2B BAT M2-Track MLVSNet V2B PTTR CMT CXTrack STNet
Builds on: M2-Track DETR MoCo PointNet++ GT-AUG
Closest papers by meaning
- M3SOT: Multi-Frame, Multi-Field, Multi-Space 3D Single Object Tracking (AAAI 2024 · similarity 0.88)
- UAWTrack: Universal 3D Single Object Tracking in Adverse Weather (AAAI 2025 · similarity 0.88)
- Towards Category Unification of 3D Single Object Tracking on Point Clouds (ICLR 2024 · similarity 0.87)
- PTT: Point-Trajectory Transformer for Efficient Temporal 3D Object Detection (CVPR 2024 · similarity 0.86)
- Motion Cues from Image-based Point Tracking for LiDAR Scene Flow Estimation (NeurIPS 2026 · similarity 0.85)
- SeMoLi: What Moves Together Belongs Together (CVPR 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.