Research map / 3D vision / Gaussian splatting
DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input
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 DrivingForward, a feed-forward 3DGS model that reconstructs surround-view driving scenes in real time from flexible multi-frame inputs, trained self-supervised without depth GT or extrinsics.
Problem
Real-time, generalizable 3D reconstruction of driving scenes from sparse surround-view cameras with minimal overlap, variable input frame counts, and no reliance on depth ground truth or world-to-camera extrinsics.
Approach
Predict Gaussian primitives per pixel from each image independently: a self-supervised pose and depth network (photometric loss over temporal, spatial and spatial-temporal contexts) localizes primitives with scale-aware depth, and a Gaussian network predicts covariance, opacity and spherical harmonics from depth and shared image features, enabling flexible multi-frame input.
Key results
- In multi-frame mode on nuScenes (352x640), the method reaches 26.06 PSNR / 0.781 SSIM / 0.215 LPIPS, versus 25.00 PSNR for pixelSplat and 22.83 for MVSplat.
- In single-frame mode at 114x228, it reaches 21.76 PSNR / 0.767 SSIM, above DistillNeRF's 20.78 PSNR.
- Against scene-optimized 3DGS in single-frame mode, it reaches 21.84 PSNR in 0.29 s per scene versus 19.57 PSNR for 3DGS at about 9 min per scene.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DrivingForward instead of pixelSplat | nuScenes MF reconstruction · nuScenes MF, 352 × 640 | PSNR: 25.00 -> 26.06 | Helped |
| Use DrivingForward instead of pixelSplat | nuScenes SF reconstruction · nuScenes SF, 352 × 640 | PSNR: 17.54 -> 21.67 | Helped |
| Use DrivingForward instead of 3DGS | nuScenes scene reconstruction · nuScenes SF, three validation scenes | PSNR: 19.57 -> 21.84 | Helped |
| Use DrivingForward instead of 3DGS | nuScenes scene reconstruction · nuScenes MF, three validation scenes | PSNR: 25.87 -> 26.66 | Helped |
| Add joint end-to-end training (ablation) | DrivingForward · nuScenes SF, 352 × 640 | PSNR: 21.50 -> 21.67 | Helped |
| Add shared image features from the depth network (ablation) | DrivingForward · nuScenes SF, 352 × 640 | PSNR: 19.48 -> 21.67 | Helped |
| Use scale-aware localization instead of monocular depth estimation (ablation) | DrivingForward · nuScenes SF, 352 × 640 | PSNR: 11.96 -> 21.67 | Helped |
Surprising findings
- Naively applying pixelSplat and MVSplat in single-frame mode yields very poor results (about 17.5 PSNR), since they need dense overlap.
- Replacing scale-aware localization with a monocular depth model drops PSNR from 21.67 to 11.96.
Limitations
- DistillNeRF comparison uses its reported numbers because its code is not released, and baselines were retrained on the authors' data loader.
- Evaluation relies on novel-view synthesis metrics only, with no downstream driving-task evaluation.
- Scene-optimized comparison uses only three validation scenes.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: PSNR SSIM LPIPS test time per scene training and inference time GPU memory
Baselines: MVSplat pixelSplat DistillNeRF EmerNeRF UniPAD SelfOcc 3DGS
Builds on: 3D Gaussian Splatting self-supervised surround-view depth estimation (SurroundDepth) pixelSplat MVSplat
Closest papers by meaning
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.92)
- DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes (CVPR 2025 · similarity 0.90)
- DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving (NeurIPS 2025 · similarity 0.89)
- FlexDrive: Toward Trajectory Flexibility in Driving Scene Gaussian Splatting Reconstruction and Rendering (CVPR 2025 · similarity 0.89)
- FRUC: Feedforward Dynamic Scene Reconstruction from Uncalibrated Collaborative Driving Views (NeurIPS 2026 · similarity 0.89)
- Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting (NeurIPS 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.