Research map / Point clouds / Point cloud classification
ZeroFlow: Scalable Scene Flow via Distillation
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 Scene Flow via Distillation (SFvD) and its instance ZeroFlow, which distills pseudo-labels from the optimization-based NSFP teacher into a fast feedforward FastFlow3D student, achieving state-of-the-art scene flow on Argoverse 2 with zero human labels.
Problem
Scene flow estimation on full-size LiDAR point clouds is either accurate but too slow (test-time optimization taking tens of seconds) or fast but requires costly human annotations for feedforward models.
Approach
Use a label-free optimization method (NSFP) to generate pseudo-labels on large, diverse unlabeled point cloud pairs, then supervise an efficient feedforward network (FastFlow3D, with an XL variant) on them, so the student is fast at test time and benefits from data scale.
Key results
- ZeroFlow XL 3X achieves Threeway EPE 0.054 on Argoverse 2 validation, versus 0.055 for supervised FastFlow3 XL and 0.055 for Chodosh et al. (optimization-based), with zero human labels.
- ZeroFlow 1X runs at 29.33 ms per full-size point cloud versus roughly 35,281 ms for Chodosh et al., over 1000x faster.
- Pseudo-labeling the Argoverse 2 train split with NSFP cost about $394 versus an estimated $787,500 for human annotation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train ZeroFlow 1X on additional unlabeled Argoverse 2 LiDAR data (3X) (ablation) | ZeroFlow · Argoverse 2 Sensor val | Threeway EPE: 0.088 -> 0.064 | Helped |
| Increase the dataset scale to 5X (ablation) | ZeroFlow · Argoverse 2 Sensor val | Threeway EPE: 0.088 -> 0.056 | Helped |
| Use the XL student backbone and additional unlabeled data (ablation) | ZeroFlow · Argoverse 2 Sensor val | Threeway EPE: 0.070 -> 0.054 | Helped |
| Train on diverse AV2 LiDAR subset data instead of AV2 Sensor data (ablation) | ZeroFlow 1X · Argoverse 2 Sensor val | Threeway EPE: $0.088$ -> $0.082$ | Helped |
| Replace NSFP pseudo-labels with Chodosh-refined pseudo-labels (ablation) | ZeroFlow 1X · Argoverse 2 Sensor val | Threeway EPE: $0.088$ -> $0.085$ | No clear effect |
| Replace NSFP pseudo-labels with TruncatedChamfer pseudo-labels (ablation) | ZeroFlow 1X · Argoverse 2 Sensor val | Threeway EPE: $0.088$ -> $0.105$ | Hurt |
| Apply speed-based point weighting instead of uniform weighting (ablation) | ZeroFlow 1X · Argoverse 2 Sensor val | Threeway EPE: $0.088$ -> $0.084$ | Helped |
| Apply speed-based point weighting instead of uniform weighting (ablation) | ZeroFlow XL 3X · Argoverse 2 Sensor val | Threeway EPE: $0.053$ -> $0.056$ | Hurt |
| Replace semantic point weighting with uniform weighting (ablation) | FastFlow3D · Argoverse 2 Sensor val | Threeway EPE: $0.071$ -> $0.081$ | Hurt |
| Replace semantic point weighting with speed-based weighting (ablation) | FastFlow3D · Argoverse 2 Sensor val | Threeway EPE: $0.071$ -> $0.081$ | Hurt |
| Distill NSFP pseudo-labels into a feedforward model | ZeroFlow 1X · Argoverse 2 Sensor val | Threeway EPE: 0.071 -> 0.088 | Hurt |
| Distill NSFP pseudo-labels into a feedforward model | ZeroFlow 1X · Waymo Open | Threeway EPE: 0.078 -> 0.092 | Hurt |
Surprising findings
- A student trained on imperfect NSFP pseudo-labels (ZeroFlow 1X, Threeway EPE 0.088) is slightly worse than human-supervised FastFlow3D (0.071) at equal data, but scaling unlabeled data to 3X surpasses both.
- Using a better teacher (Chodosh refined labels) gave no meaningful improvement over NSFP pseudo-labels.
Limitations
- ZeroFlow inherits the biases and failure modes of its NSFP pseudo-labels.
- Teacher quality improvements (Chodosh) did not translate into student gains.
- Speed-based point weighting helped 1X but hurt 3X scaling.
Experimental setup
Datasets: Argoverse 2 Sensor Argoverse 2 LiDAR Waymo Open
Benchmarks: Argoverse 2 Self-Supervised Scene Flow Challenge Argoverse 2 Sensor validation Waymo Open validation
Metrics: Threeway EPE Dynamic FG EPE Static FG EPE Static BG EPE runtime (ms)
Baselines: FastFlow3D NSFP w/ Motion Compensation Chodosh et al. ICP Gojcic Sim2Real EgoFlow PPWC FlowStep3D FlowNet3D
Builds on: Neural Scene Flow Prior FastFlow3D PointPillars
Closest papers by meaning
- Zero-Shot Monocular Scene Flow Estimation in the Wild (CVPR 2025 · similarity 0.89)
- DeltaFlow: An Efficient Multi-frame Scene Flow Estimation Method (NeurIPS 2025 · similarity 0.88)
- 3DSFLabelling: Boosting 3D Scene Flow Estimation by Pseudo Auto-labelling (CVPR 2024 · similarity 0.88)
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.88)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.87)
- VoteFlow: Enforcing Local Rigidity in Self-Supervised Scene Flow (CVPR 2025 · 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.