MVSplat is an innovative model that efficiently predicts clean feed-forward 3D Gaussians from sparse multi-view images. By using a cost volume representation via plane sweeping, MVSplat accurately localizes Gaussian centers and provides valuable geometry cues for depth estimation. Impressively, MVSplat outperforms the state-of-the-art pixelSplat method with 10x fewer parameters, faster inference speed (22 fps), and better appearance and geometry quality. Additionally, MVSplat excels in cross-dataset generalization, showcasing its ability to generalize to novel scenes. The research, supported by Monash FIT Start-Up Grant and EPSRC SYN3D, is documented in an article by Chen et al. in arXiv.
https://donydchen.github.io/mvsplat/