Splatt3r-SfM
N-view Gaussian splats from unposed images, with no per-scene optimisation
A 3D Vision course project that couples two feed-forward models that each stop one step short of what you want:
- MASt3R-SfM recovers N-view camera poses from an unposed image collection.
- Splatt3r predicts a Gaussian splat, but only from an image pair.
Splatt3r-SfM fuses the pairwise splats into a single aligned scene using MASt3R-SfM’s poses, giving N-view novel view synthesis from unposed images in a single forward pass — no per-scene optimisation, unlike classical Gaussian splatting.