Research map / 3D vision / Depth and stereo
DifFlow3D: Toward Robust Uncertainty-Aware Scene Flow Estimation with Iterative Diffusion-Based Refinement
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 DifFlow3D, a plug-and-play iterative diffusion-based refinement module with per-point uncertainty estimation for point-cloud scene flow, trained on FlyingThings3D and evaluated on FlyingThings3D and KITTI.
Problem
Scene flow estimation from point clouds suffers from unreliable point correlation (locally constrained KNN search, matching noise) and accumulated error in coarse-to-fine refinement; existing methods also lack per-point reliability estimates.
Approach
Reformulate coarse-to-fine flow refinement as a conditional diffusion model that denoises flow residuals conditioned on geometry features, cross-frame cost volume and coarse flow embeddings, jointly predicting a per-point uncertainty that narrows during training.
Key results
- Reduces EPE3D by 24.0% on FlyingThings3D (0.0114 vs MSBRN 0.0150) and 29.1% on KITTI (0.0078 vs 0.0110), trained only on FlyingThings3D.
- As a plug-and-play module on MSBRN, it cuts KITTI EPE3D to 0.0078 m, reported as millimeter-level accuracy.
- Adds only about 3.3% runtime overhead over MSBRN (228.3 ms vs 221.1 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add diffusion-based refinement to the coarse initialization (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.1559 -> 0.0114 | Helped |
| Use a GRU-based rather than transformer-based denoising network (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0374 -> 0.0114 | Helped |
| Use a GRU-based rather than MLP-based denoising network (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0244 -> 0.0114 | Helped |
| Add jointly estimated uncertainty (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0147 -> 0.0114 | Helped |
| Use binary rather than continuous uncertainty ground truth (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0178 -> 0.0114 | Helped |
| Estimate uncertainty jointly rather than separately from scene flow (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0170 -> 0.0114 | Helped |
| Add cost volume as a condition (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0621 -> 0.0114 | Helped |
| Add coarse flow embedding as a condition (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0245 -> 0.0114 | Helped |
| Add geometry feature as a condition (ablation) | DifFlow3D · FlyingThings3D | EPE3D: 0.0150 -> 0.0114 | Helped |
| Add DSFR to MSBRN | MSBRN · FlyingThings3D and KITTI | EPE3D: 0.0150 -> 0.0114; 0.0110 -> 0.0078 | Helped |
| Add DSFR to 3DFlow | 3DFlow · FlyingThings3D and KITTI | EPE3D: 0.0281 -> 0.0242; 0.0309 -> 0.0251 | Helped |
| Add DSFR to BPF | BPF · FlyingThings3D and KITTI | EPE3D: 0.0280 -> 0.0247; 0.0300 -> 0.0265 | Helped |
Surprising findings
- Removing the cost volume condition raises EPE3D by about 4.5x (0.0621 vs 0.0114), showing strong dependence on correlation conditioning.
- Predicting uncertainty and flow with separate heads instead of jointly degrades EPE3D noticeably (0.0170 vs 0.0114).
Limitations
- Diffusion sampling with T=1000 steps (DDIM) and iterative refinement layers may be costly, though overhead vs MSBRN is small.
- Uncertainty ground truth depends on hand-tuned decaying thresholds E1 and E2 derived from ground-truth errors.
- Dynamic objects such as moving cars still have larger uncertainty intervals, indicating remaining difficulty.
Experimental setup
Datasets: FlyingThings3D KITTI scene flow
Benchmarks: FlyingThings3D KITTI scene flow
Metrics: EPE3D Acc3DS Acc3DR Outliers EPE2D Acc2D
Baselines: FlowNet3D HPLFlowNet PointPWC-Net HALFlow FLOT HCRF-Flow PV-RAFT FlowStep3D RCP 3DFlow BPF PT-FlowNet
Builds on: MSBRN 3DFlow FlowNet3D DDIM PointNet++
Closest papers by meaning
- DiffSF: Diffusion Models for Scene Flow Estimation (NeurIPS 2024 · similarity 0.91)
- Injecting Frame-Event Complementary Fusion into Diffusion for Optical Flow in Challenging Scenes (NeurIPS 2025 · similarity 0.88)
- Coherent 3D Scene Diffusion From a Single RGB Image (NeurIPS 2024 · similarity 0.88)
- MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting (NeurIPS 2026 · similarity 0.87)
- DeltaFlow: An Efficient Multi-frame Scene Flow Estimation Method (NeurIPS 2025 · similarity 0.87)
- Zero-Shot Monocular Scene Flow Estimation in the Wild (CVPR 2025 · similarity 0.87)
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.