Structure-from-Motion (SfM) is a common but cumbersome step in 3D reconstruction, used to estimate camera poses before building a scene. Pose-free 3D Gaussian Splatting (3DGS) aims to skip this step entirely by reconstructing scenes directly from unposed image sequences.
CrossFit-GS, introduced in a new arXiv paper, tackles this problem with a deferred-commitment progressive strategy. The method incorporates out-of-fold pose estimation, allowing the system to refine poses while building the 3D representation.
The approach could simplify 3D reconstruction pipelines, making it easier to generate 3D scenes from casual image collections without the overhead of SfM preprocessing.