Reproduce “TinySplat: Feedforward Approach for Generating Compact 3D Scene Representation” 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
Zetian Song, Jiaye Fu, Jiaqi Zhang, Xiaohan Lu, Chuanmin Jia, Siwei Ma, Wen Gao
TinySplat introduces a feedforward, training-free compression framework for 3D Gaussian Splatting models, achieving over 100x data reduction while preserving high rendering fidelity. The method significantly outperforms existing state-of-the-art training-free approaches by reducing encoding time by 75% and decoding time by over 100x.
Key insights
- Transforming world-space geometric parameters into camera space through View-Projection Transform significantly improves their regularity and local correlation, making them more efficiently compressible.
- Many Spherical Harmonic (SH) basis functions correspond to unobserved viewing directions; adapting SH bases based on expected visibility allows for substantial perceptual redundancy reduction.
- By treating transformed Gaussian feature maps as 2D images and employing standard video codecs, significant spatial redundancy can be effectively exploited in a training-free manner.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.