Reproduce “SplArt: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting” 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
Shengjie Lin, Jiading Fang, Muhammad Zubair Irshad, Vitor Campagnolo Guizilini, Rares Andrei Ambrus, Greg Shakhnarovich, Matthew R. Walter
SPLART, developed by researchers from Toyota Technological Institute at Chicago and Toyota Research Institute, presents a framework for part-level reconstruction and articulation estimation of objects using 3D Gaussian Splatting from just two sets of posed RGB images. This approach achieved a 99.0% success rate on the PARIS-PMS dataset with substantially lower articulation errors than baselines, concurrently enabling real-time photorealistic rendering for novel views and articulation states.
Key insights
- Augmenting 3DGS with a differentiable mobility parameter enables precise part segmentation and kinematic inference through gradient-based optimization.
- A multi-stage, decoupled optimization process effectively addresses the non-convexity of joint reconstruction and articulation problems, leading to improved robustness and accuracy.
- Geometric consistency losses, specifically the cross-static and cross-mobile formulations, provide a powerful self-supervision mechanism for accurately estimating articulation parameters from minimal 2D image inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.