Reproduce “PocketSplat: Mobile Gaussian Reconstruction via World-Space Latent Allocatio” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenzhi Guo, Xianda Chen, Dongxuan Chen, Guangchi Fang, Bing Wang
PocketSplat reconstructs mobile 3D scenes with a requested Gaussian count. It groups image evidence by predicted world-space location, allocates capacity across cells, pools retained evidence before decoding, and adapts covariance support. On Mip-NeRF 360, PocketSplat Light uses 75.4% fewer Gaussians than MVSplat-Stream and scores 20.018 dB PSNR versus 17.875; this modified comparator runs on the phone, unlike native MVSplat and DepthSplat.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.