MrNeRF (@janusch_patas)
2024-11-12 | โค๏ธ 130 | ๐ 24
SplatFormer: Point Transformer for Robust 3D Gaussian Splatting
Contributions: โข We introduce OOD-NVS, a new experimental protocol specifically designed to evaluate the performance of NVS methods when rendering 3D scenes from novel viewing angles that fall outside the distribution of input views. Our results demonstrate that existing methods struggle to generalize under the OOD-NVS protocol;
โข We propose SplatFormer, a novel learning-based model that refines flawed 3D Gaussian splats to mitigate artifacts in OOD views. SplatFormer is the first approach to apply the point transformer to 3DGS processing, effectively leveraging multi-view information from a dense set of input views and learning a 3D rendering prior to remove artifacts;
โข We demonstrate that SplatFormer significantly improves the performance of 3DGS-based methods on OOD-NVS tasks, achieving substantial gains in object-centric scenes, while also demonstrating potential for application in unbounded environments.
๋ฏธ๋์ด
