Research map / Detection and segmentation / Open-set and few-shot detection
Floxels: Fast Unsupervised Voxel Based 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 Floxels, a test-time optimization method for scene flow that replaces the MLP with a voxel grid, uses a multi-scan distance transform loss and a cluster consistency loss, achieving near-EulerFlow quality at a fraction of runtime.
Problem
Unsupervised scene flow estimation from lidar point clouds: existing test-time optimization methods are either slow (EulerFlow, ~24 h per sequence) or low quality (NSFP, FNSF), with windmill artifacts, shadow errors and false associations.
Approach
Replace the coordinate MLP in neural scene flow prior with an explicit trilinear-interpolated voxel grid of 3D flow vectors, optimize flow jointly over multiple scans using distance transform losses with time-decay weights, and add a DBSCAN-based cluster consistency loss plus flow norm regularizer.
Key results
- On the Argoverse 2 challenge test set, Floxels 5 reaches mdnEPE 0.1676 (13 scans: 0.1538), second only to EulerFlow (0.1304) among unsupervised methods.
- Floxels runs in about 9 minutes per 156-scan sequence versus ~24 hours for EulerFlow, a speedup of roughly 60-140x; it is ~14.4x faster than NSFP on larger point clouds.
- On nuScenes-mini ablation, removing the cluster loss raises dynamic EPE from 0.085 to 0.201, showing the cluster loss is critical.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the MLP with a voxel grid | Floxels · Argoverse 1 test, dynamic points | EPE: 0.200 -> 0.104 | Helped |
| Add cluster consistency loss (ablation) | Floxels · nuScenes mini, dynamic points | EPE: 0.201 -> 0.085 | Helped |
| Add cluster consistency loss (ablation) | Floxels · nuScenes mini, dynamic points | Acc5: 0.133 -> 0.537 | Helped |
| Add flow norm loss (ablation) | Floxels · nuScenes mini, dynamic points | Acc5: 0.528 -> 0.537 | Helped |
| Add flow norm loss (ablation) | Floxels · nuScenes mini, dynamic points | EPE: 0.084 -> 0.085 | No clear effect |
| Use 3 scans instead of the 5-scan setting (ablation) | Floxels · nuScenes mini, dynamic points | EPE: 0.085 -> 0.095 | Hurt |
| Use 7 scans instead of the 5-scan setting (ablation) | Floxels · nuScenes mini, dynamic points | EPE: 0.085 -> 0.082 | Helped |
| Use 9 scans instead of the 5-scan setting (ablation) | Floxels · nuScenes mini, dynamic points | EPE: 0.085 -> 0.078 | Helped |
| Use 11 scans instead of the 5-scan setting (ablation) | Floxels · nuScenes mini, dynamic points | EPE: 0.085 -> 0.076 | Helped |
| Increase voxel size from 0.5 m to 2.0 m (ablation) | Floxels · Argoverse 2 validation | mdnEPE: 0.217 -> 0.229 | Hurt |
| Increase voxel size from 0.5 m to 2.0 m (ablation) | Floxels · Argoverse 2 validation | Time (s): 3.81 -> 3.41 | Helped |
| Use a voxel grid instead of an MLP | Floxels · Argoverse 1 test, dynamic points | Time (s): 63.01 -> 4.38 | Helped |
Surprising findings
- The MLP-based neural prior produces windmill artifacts (false flow in empty regions) that inflate metrics because the loss only measures points, so MLP-based methods appear better than they are.
- Training NSFP without early stopping or with a 16-layer network yields no or only mild improvements, suggesting the poor dynamic performance is not due to undertraining or capacity.
Limitations
- Voxel size trades spatial resolution against runtime and memory; large voxels limit flow detail, small voxels slow optimization.
- Still second to EulerFlow in accuracy on the Argoverse 2 challenge, which is ~60-140x slower.
- Relies on ego-motion corrected data and on a constant-velocity assumption across adjacent scans; hyperparameters were tuned on nuScenes-mini.
Experimental setup
Datasets: Argoverse 2 Argoverse 1 nuScenes nuScenes-mini
Benchmarks: Argoverse 2 Scene Flow Challenge test set Argoverse 1 test set nuScenes validation Argoverse 2 validation
Metrics: mean dynamic normalized EPE dynamic normalized EPE EPE Acc5 Acc10 angle error runtime
Baselines: NSFP FNSF EulerFlow DifFlow3D Flow4D DeFlow SeFlow ICP-Flow ZeroFlow TrackFlow
Builds on: Neural Scene Flow Prior (NSFP) Fast Neural Scene Flow (FNSF) Plenoxels / voxel grid radiance fields EulerFlow DBSCAN
Closest papers by meaning
- VoxelSplat: Dynamic Gaussian Splatting as an Effective Loss for Occupancy and Flow Prediction (CVPR 2025 · similarity 0.87)
- DiffSF: Diffusion Models for Scene Flow Estimation (NeurIPS 2024 · similarity 0.87)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.86)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.86)
- DifFlow3D: Toward Robust Uncertainty-Aware Scene Flow Estimation with Iterative Diffusion-Based Refinement (CVPR 2024 · similarity 0.86)
- MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting (NeurIPS 2026 · 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.