Reproduce “ArtSplat: Feed-Forward Articulated 3D Gaussian Splatting from Sparse Multi-State Uncalibrated Views” 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

Inseo Lee, Yoonji Kim, Eugene Sohn, Jiwoong Lee, Jungmin You, Joonseok Lee, Jin-Hwa Kim

ArtSplat provides a feed-forward framework for articulated 3D Gaussian Splatting, efficiently reconstructing object geometry and joint parameters from sparse, multi-state, uncalibrated RGB views. The method reduces per-object reconstruction time to under 2 seconds while achieving improved accuracy in joint axis estimation and geometry over existing methods.

Key insights

  • Explicitly modeling inter-state differences via Cross-State Attention and state tokens is critical for inferring accurate joint parameters from sparse, multi-state observations.
  • A dense, per-pixel joint map representation allows for end-to-end differentiable estimation of articulation parameters directly within a feed-forward model.
  • A two-stage training strategy enhances stability by first optimizing geometry and joint parameters with direct supervision, then refining appearance with photometric losses, which mitigates instability from early noisy joint predictions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.