Research map / 3D vision / Gaussian splatting
SplatFlow: Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving
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 SplatFlow, a self-supervised dynamic Gaussian splatting method that uses Neural Motion Flow Fields to decompose static and dynamic urban scenes and render RGB, depth, and flow without 3D bounding boxes.
Problem
Dynamic urban scene reconstruction and novel view synthesis with Gaussian splatting typically requires expensive 3D bounding box and tracking annotations, limiting scalability.
Approach
Models temporal motion of LiDAR points and 4D Gaussians as continuous Neural Motion Flow Fields, pretrained with bidirectional Chamfer distance on LiDAR and jointly optimized with Gaussians, plus optical flow distillation from SEA-RAFT to separate static and dynamic elements without object labels.
Key results
- On Waymo image reconstruction, SplatFlow reaches 33.64 PSNR / 0.951 SSIM / 0.198 LPIPS versus PVG's 32.46 / 0.910 / 0.373.
- On Waymo novel view synthesis, SplatFlow reaches 28.71 PSNR / 0.874 SSIM / 0.239 LPIPS versus PVG's 28.11 / 0.849 / 0.279.
- On KITTI novel view synthesis, SplatFlow reaches 28.32 PSNR / 0.932 SSIM / 0.089 LPIPS versus PVG's 27.43 / 0.896 / 0.114.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LiDAR-pretrained NMFF prior (ablation) | SplatFlow · Waymo, novel view synthesis | PSNR: 27.69 -> 28.99 | Helped |
| Add NMFF optimization with 4D Gaussians on image data (ablation) | SplatFlow · Waymo, novel view synthesis | PSNR: 28.14 -> 28.99 | Helped |
| Add optical flow distillation (ablation) | SplatFlow · Waymo, novel view synthesis | PSNR: 28.28 -> 28.99 | Helped |
| Add LiDAR-pretrained NMFF prior (ablation) | SplatFlow · Waymo dynamic objects, novel view synthesis | PSNR: 27.28 -> 28.90 | Helped |
| Add NMFF optimization with 4D Gaussians on image data (ablation) | SplatFlow · Waymo dynamic objects, novel view synthesis | PSNR: 27.97 -> 28.90 | Helped |
| Add optical flow distillation (ablation) | SplatFlow · Waymo dynamic objects, novel view synthesis | PSNR: 28.51 -> 28.90 | Helped |
| Replace PVG with SplatFlow | Dynamic scene reconstruction · Waymo, image reconstruction | PSNR: 32.46 -> 33.64 | Helped |
| Replace PVG with SplatFlow | Dynamic scene reconstruction · KITTI, novel view synthesis | PSNR: 27.43 -> 28.32 | Helped |
Limitations
- Relies on pseudo ground truth optical flow from SEA-RAFT and sky masks from SegFormer during training.
- Dynamic object masks for PSNR* are derived from 2D ground-truth bounding boxes following EmerNeRF.
Experimental setup
Datasets: Waymo Open Dataset KITTI
Benchmarks: Waymo Open Dataset KITTI
Metrics: PSNR SSIM LPIPS
Baselines: S-NeRF StreetSurf 3DGS NSG Mars SUDS EmerNeRF PVG StreetGaussian NeRF
Builds on: 3D Gaussian Splatting PVG EmerNeRF SEA-RAFT
Closest papers by meaning
- DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes (CVPR 2025 · similarity 0.90)
- Gaussian Splatting Lucas-Kanade (ICLR 2025 · similarity 0.90)
- HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene (NeurIPS 2025 · similarity 0.89)
- SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.88)
- H3D-DGS: Exploring Heterogeneous 3D Motion Representation for Deformable 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.88)
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.