Research map / Point clouds / Point cloud classification
Motion Cues from Image-based Point Tracking for LiDAR Scene Flow Estimation
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 TrackCue, which uses image point tracking (CoTracker3) with ego-motion compensation and lifting to refine DUFOMap dynamic pseudo-labels, improving self-supervised LiDAR scene flow for SeFlow, SeFlow++, and TeFlow.
Problem
Self-supervised LiDAR scene flow relies on static-dynamic labels from LiDAR-only ray-casting (DUFOMap), which are noisy under sparsity and occlusion, degrading flow learning.
Approach
Track LiDAR-projected pixels with an image point tracker, compare tracks with ego-induced rigid trajectories to isolate true object motion, and lift the resulting moving trajectories back to LiDAR points via nearest-neighbor association to refine dynamic labels.
Key results
- On the AV2 leaderboard, TrackCue reduces overall Dynamic Bucket-Normalized EPE of SeFlow++ from 0.7025 to 0.6619 (4.06 points).
- On the dynamic auto-labeling comparison with SeFlow, TrackCue raises F1 from 36.74% to 42.37% and precision from 24.62% to 31.38% versus DUFOMap.
- On the AV2 validation set, TrackCue reduces SeFlow++ mean three-way EPE from 4.78 to 4.29 cm.
Surprising findings
- TrackCue with TAPNext yields highest precision but much lower recall, which hurts downstream scene flow.
- Gains on the in-domain Argoverse 2 split are modest, while improvements on unseen datasets such as AEVAScenes are much larger.
Limitations
- Projection of 3D points to 2D loses geometry, so small or distant objects may be missed or over-suppressed.
- Scope limited to pseudo-label generation rather than joint cross-modal feature learning.
- Performance depends on camera quality, calibration, and point tracker failures.
Experimental setup
Datasets: Argoverse 2 Waymo nuScenes TruckScenes AEVAScenes
Benchmarks: Argoverse 2 validation Argoverse 2 scene flow leaderboard (AV2, Waymo, nuScenes, TruckScenes, AEVAScenes)
Metrics: Three-way EPE Dynamic Bucket-Normalized EPE precision recall F1 score accuracy
Baselines: SeFlow SeFlow++ TeFlow DUFOMap NSFP FastNSF ICP-Flow ZeroFlow VoteFlow
Builds on: SeFlow SeFlow++ TeFlow DUFOMap CoTracker3
Closest papers by meaning
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.89)
- 3DSFLabelling: Boosting 3D Scene Flow Estimation by Pseudo Auto-labelling (CVPR 2024 · similarity 0.89)
- SeMoLi: What Moves Together Belongs Together (CVPR 2024 · similarity 0.87)
- Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations (CVPR 2024 · similarity 0.86)
- VoteFlow: Enforcing Local Rigidity in Self-Supervised Scene Flow (CVPR 2025 · similarity 0.86)
- ZeroFlow: Scalable Scene Flow via Distillation (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.