No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images

NoPoSplat Novel View Synthesis 15 ms Pose Estimation Downstream Tasks 3D Gaussians in Canonical Space Unposed Images Figure 1: NoPoSplat. Given sparse unposed images, our method reconstructs 3D Gaussians of different views in a canonical space using a feed-forward network. The resulting 3D Gaussians can be utilized for accurate relative camera pose estimation and high-quality novel view synthesis. The input images for illustration are extracted from a Sora-generated video.