Reproduce “AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling” 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
Yingji Zhong, Dave Zhenyu Chen, Fuzhao Ou, Youyu Chen, Zhihao Li, Lanqing Hong, Dan Xu
AsySplat predicts renderable 3D Gaussians from posed images. It uses coarse tokens across views to estimate geometry and fine tokens within each frame for appearance, with information exchanged between branches. On DL3DV-140, using 32 views at 540×960, it reaches 24.00 PSNR in under a second versus 23.60 for 3DGS after 30,000 optimization iterations; this compares feed-forward prediction with per-scene optimization.
Key insights
- High-precision geometry is not strictly required for high-quality Novel View Synthesis (NVS) in 3DGS, suggesting computational resources can be reallocated away from overly fine geometric detail.
- Modeling high-precision appearance attributes is generally easier than geometry recovery, particularly once a foundational multi-view depth is established, and can be achieved with less intensive inter-frame computations.
- An asymmetric allocation of computational effort and parameters, tailored to the distinct demands of geometry and appearance modeling, can lead to substantial efficiency gains without compromising NVS performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.