GenRe: Diffusion-Guided Enhancer for Urban Scene Reconstruction
A new tool named GenRe has been developed by researchers, serving as a diffusion-guided enhancer for reconstructing urban scenes. While existing neural rendering techniques provide high-quality rendering along specific paths, they struggle with significant viewpoint changes. GenRe can enhance any pretrained 3D Gaussian model in just minutes, allowing for generalization beyond restricted synthesized perspectives without the need for expensive scene-specific optimization. This innovation tackles the shortcomings of current diffusion-based enhancements that necessitate per-scene adjustments and lack generalizability. Its implications are significant for advancements in self-driving technology and closed-loop simulations.
Key facts
- GenRe is a diffusion-guided generalizable enhancer for urban scene reconstruction.
- Current neural rendering approaches degrade under large viewpoint shifts.
- GenRe takes any pretrained 3D Gaussian representation as input.
- GenRe fixes deficiencies within a few minutes.
- Existing diffusion-based enhancements require costly per-scene optimization.
- Distilled representations from existing methods fail to generalize beyond limited synthesized views.
- Urban scene reconstruction is used for self-driving development and testing.
- The method aims to improve closed-loop simulation applicability.
Entities
Institutions
- arXiv