Projects

InstantSplat++: Registration-Free Gaussian Splatting from Sparse Image Sequences

Gonzalo Lope Carrasco, Björn Kremser, Moritz Pötzsch, Osama Salem

Course project, Technical University of Munich ·

Supervisor Prof. Angela Dai

An unposed image sequence is turned by InstantSplat++ into a Gaussian scene, rendered as novel views, depth, normals and surfels.
Overview. Starting from an unposed image sequence, InstantSplat++ reconstructs a Gaussian scene that can be rendered as novel views, depth and normals.

TL;DR

Problem
Sparse-view Gaussian splatting usually depends on structure-from-motion, which is slow and unreliable with few views. InstantSplat drops SfM but still relies on pairwise predictions and global alignment, which over-parameterizes the scene.
Approach
Spann3R predicts a global point cloud in one pass; a shared focal length is estimated with the Weiszfeld algorithm and poses with PnP-RANSAC. Cumulative masking (frustum overlap and depth consistency) removes redundant points, then Gaussians and camera poses are optimized jointly — for both 3DGS and 2DGS.
Contribution
On ScanNet++ with 3–15 views, InstantSplat++ achieves higher PSNR than InstantSplat in every setting (e.g. 38.07 vs 30.70 dB with 3DGS at 9 views), with 3.7–9.8× faster initialization.