Projects

I-GaSp3R: Incremental Gaussian Splatting for 3D Reconstruction

Tobias Steinle, Osama Salem

ADL4CV course project, Technical University of Munich ·

Supervisors Lukas Höllein, Prof. Matthias Nießner

A stack of unposed input images is processed by I-GaSp3R into a global 3D Gaussian reconstruction, which a rasterizer renders as novel views.
Overview. Given unposed images of variable length, I-GaSp3R incrementally reconstructs 3D Gaussians in a single feed-forward pass, without pre-processing or pose estimation.

TL;DR

Problem
Most 3D reconstruction and novel-view-synthesis methods need calibrated cameras and offline optimization, which limits them in dynamic or unstructured settings.
Approach
Combine the spatial memory of Spann3R with the Gaussian head of Splatt3R to predict 3D Gaussians frame by frame. Training uses a masked photometric loss, a temporal-consistency loss across consecutive renderings, confidence weighting and frequency-based gradient normalization.
Contribution
Real-time, incremental 3D Gaussian splatting at 21 FPS on a single RTX 3080. On ScanNet++ it reaches 26.46 / 24.90 / 20.81 dB PSNR with 2 / 10 / 50 views (baselines as reported by PREF3R on a similar, not identical, split).