Research map / Label-efficient and robust learning / Self-supervised and contrastive
VoteFlow: Enforcing Local Rigidity in Self-Supervised Scene Flow
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.
Adds a differentiable Voting Module over pillar features that votes for shared translations among nearby points, enforcing local rigidity by architecture in self-supervised LiDAR scene flow estimation.
Problem
Self-supervised scene flow from LiDAR scans typically enforces local motion rigidity only through losses, post-processing, or clustering, lacking an architectural inductive bias for it.
Approach
A lightweight differentiable voting module builds a discretized translation voting space per non-empty pillar, filled with cosine-similarity votes from neighboring pillars' features across time, so that nearby points' shared motion is identified end-to-end; operating on pillars keeps computation low.
Key results
- On Argoverse 2 test, VoteFlow improves dynamic normalized EPE mean to 0.249 vs SeFlow's 0.267 (2.0 percentage points better) among self-supervised models.
- On Waymo validation, VoteFlow trained on Argoverse 2 and tested directly reaches FD EPE 0.142 vs SeFlow's 0.155 when SeFlow is also trained on Argoverse 2.
- Voting adds marginal overhead: about 25.6 ms per sample on an A100 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Voting Module for local rigidity | VoteFlow · Argoverse 2 test split | Mean dynamic normalized EPE: 0.267 -> 0.249 | Helped |
| add the Voting Module for local rigidity | VoteFlow · Waymo Open validation split, trained on Waymo | Foreground dynamic EPE: 0.151 -> 0.117 | Helped |
| add the Voting Module for local rigidity | VoteFlow · Waymo Open validation split, cross-dataset | Foreground dynamic EPE: 0.155 -> 0.142 | Helped |
| add the Voting Module for local rigidity | VoteFlow · Waymo Open validation split, cross-dataset | Foreground static EPE: 0.018 -> 0.014 | Helped |
| increase M from 8 to 16 or 32 (ablation) | VoteFlow · Argoverse 2 validation split | Mean dynamic normalized EPE: 0.347 -> 0.338; 0.347 -> 0.337 | Helped |
| increase M from 8 to 64 (ablation) | VoteFlow · Argoverse 2 validation split | Mean dynamic normalized EPE: 0.347 -> 0.341 | Hurt |
| reduce M from 8 to 4 (ablation) | VoteFlow · Argoverse 2 validation split | Mean dynamic normalized EPE: 0.347 -> 0.338 | Helped |
| reduce N from 128 to 64 (ablation) | VoteFlow · Argoverse 2 validation split | Mean dynamic normalized EPE: 0.347 -> 0.341 | Hurt |
| increase N from 128 to 256 (ablation) | VoteFlow · Argoverse 2 validation split | Mean dynamic normalized EPE: 0.347 -> 0.334 | Helped |
| use a GRU decoder instead of the MLP decoder (ablation) | VoteFlow · Argoverse 2 validation split | Mean dynamic normalized EPE: 0.347 -> 0.344 | Hurt |
| use a GRU decoder instead of the MLP decoder (ablation) | VoteFlow · Argoverse 2 validation split | Pedestrian dynamic normalized EPE: 0.424 -> 0.475 | Helped |
| use a GRU decoder instead of the MLP decoder (ablation) | VoteFlow · Argoverse 2 validation split | Other Vehicles dynamic normalized EPE: 0.347 -> 0.374 | Hurt |
Surprising findings
- The GRU decoder variant of VoteFlow was worse than SeFlow on Other Vehicles, and the simpler four-layer MLP decoder performed better overall.
Limitations
- Voting space size grows with the maximum translation, so larger time intervals increase compute.
- Only translation is modeled; rotations are not handled by the voting space.
- Gap of about 11.5 percentage points remains to fully supervised Flow4D on Argoverse 2.
Experimental setup
Datasets: Argoverse 2 Waymo Open
Benchmarks: Argoverse 2 Scene Flow Challenge test split Waymo Open validation split
Metrics: Bucketed Normalized EPE 3-way EPE EPE latency (ms)
Baselines: NSFP FastNSF ZeroFlow ICP-Flow SeFlow Flow4D DeFlow FastFlow3D TrackFlow
Builds on: SeFlow ZeroFlow FastFlow3D Pillar Feature Net Chamfer loss
Closest papers by meaning
- RigidFlow: Self-Supervised Scene Flow Learning on Point Clouds by Local Rigidity Prior (CVPR 2022 · similarity 0.90)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.90)
- 3DSFLabelling: Boosting 3D Scene Flow Estimation by Pseudo Auto-labelling (CVPR 2024 · similarity 0.87)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.86)
- Motion Cues from Image-based Point Tracking for LiDAR Scene Flow Estimation (NeurIPS 2026 · similarity 0.86)
- Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations (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.