ArtSplat: Feed-Forward Articulated 3D Gaussian Splatting from Sparse Views
ArtSplat is an innovative feed-forward framework designed for articulated 3D Gaussian Splatting (3DGS), capable of reconstructing joint parameters and geometry from sparse multi-view images across various articulation states in one forward pass. In contrast to traditional techniques that depend on dense views, depth maps, or fixed joint types, ArtSplat employs a per-pixel joint map representation along with a Cross-State Attention (CSA) mechanism, allowing for the seamless integration of joint parameter estimation. This research has been made available on arXiv with the ID 2605.24304.
Key facts
- ArtSplat is the first feed-forward framework for articulated 3D Gaussian Splatting.
- It reconstructs geometry and joint parameters from sparse multi-view images.
- The method works across multiple articulation states in a single forward pass.
- A per-pixel joint map representation enables integration of joint estimation.
- A Cross-State Attention (CSA) mechanism is proposed for the pipeline.
- Existing methods rely on dense views, depth maps, or predefined joint types.
- The paper is available on arXiv with ID 2605.24304.
- The approach requires no per-object optimization.
Entities
Institutions
- arXiv