Research map / 3D vision / Gaussian splatting
FRUC: Feedforward Dynamic Scene Reconstruction from Uncalibrated Collaborative Driving Views
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 FRUC, a feedforward 3DGS framework that reconstructs dynamic driving scenes from uncalibrated multi-vehicle camera views using a VGGT backbone, a causal occlusion field, and zero-initialized residual denoising for cross-agent blind-spot completion.
Problem
Dynamic scene reconstruction from collaborative multi-vehicle driving views usually needs calibrated cameras, LiDAR initialization, and slow per-scene optimization, and naive cross-agent fusion corrupts the ego vehicle's visible geometry.
Approach
Treats multi-vehicle inputs as an unstructured ego-centric multi-camera system; an agent-wise causal masked attention derives an occlusion uncertainty prior from dynamic footprints, which gates zero-initialized residual denoising so collaborative features only fill blind spots while ego-visible regions are preserved.
Key results
- On V2X-Real (MF mode), FRUC reaches 25.89 PSNR / 0.883 SSIM / 0.076 LPIPS on full image, versus 21.85 PSNR for the next-best feed-forward method DGGT, at 0.77s inference.
- On UrbanIng-V2X zero-shot, FRUC reaches 25.07 PSNR full-image vs. 22.23 for DGGT.
- Ablation on V2X-Real: removing COF lowers full-image PSNR from 25.89 to 23.97 and removing auxiliary losses lowers it to 22.76.
Surprising findings
- Per-scene optimization baseline V2X-Gaussians obtains the best dynamic-only SSIM (0.994) and FRUC is tied at 0.994 on V2X-Real, while on UrbanIng-V2X fine-tuned FRUC has lower dynamic-only SSIM (0.819) than baselines.
- Naively adding collaborative views degrades single-agent ego-only quality for baselines such as DGGT (PSNR 24.80 with no collaborative views vs. 21.85 with two).
Limitations
- Struggles in severe occlusion scenarios where cross-agent semantic correlation is weak.
- Inaccurate dynamic probability maps can cause failures that compromise downstream scene editing.
Experimental setup
Datasets: V2X-Real UrbanIng-V2X OPV2V format benchmark (unified)
Benchmarks: V2X-Real UrbanIng-V2X
Metrics: PSNR SSIM LPIPS NIQE inference time
Baselines: EmerNeRF 3DGS V2X-Gaussians MVSplat DrivingForward STORM AnySplat DGGT
Builds on: VGGT 3D Gaussian splatting DINO DPT LPIPS
Closest papers by meaning
- One World, Dual Timeline: Decoupled Spatio-Temporal Gaussian Scene Graph for 4D Cooperative Driving Reconstruction (NeurIPS 2026 · similarity 0.90)
- DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input (AAAI 2025 · similarity 0.89)
- MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction (CVPR 2025 · similarity 0.87)
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.87)
- FreeSim: Toward Free-viewpoint Camera Simulation in Driving Scenes (CVPR 2025 · similarity 0.87)
- Flux4D: Flow-based Unsupervised 4D Reconstruction (NeurIPS 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.