VEPHand: View-Efficient Photometric Hand Performance Capture at Scale
Two new research papers published on arXiv introduce advanced techniques for capturing high-fidelity 4D hand-object interactions. The first paper, "High-Fidelity 4D Hand-Object Capture via Multi-View Spatiotemporal Tracking and Physics-Aware Gaussians," proposes a system that uses a transformer model for pose initialization and a physics-aware Gaussian optimization framework for refinement, eliminating the need for object templates or markers. The second paper, "VEPHand: View-Efficient Photometric Hand Performance Capture at Scale," presents an end-to-end pipeline for dynamic hand capture using a mask-free neural method and a physics-inspired framework for registration, designed for view-efficient setups and capable of handling intricate interactions. AI
IMPACT These papers advance the state-of-the-art in 4D reconstruction, potentially enabling more realistic virtual interactions and asset generation for embodied AI and spatial computing applications.
- Hugging Face
- arXiv
- DagsHub
- alphaXiv
- ScienceCast
- CatalyzeX
- Gotit.pub
- Embodied Ai
- spatial computing
- 4D Hand-Object Capture
- Multi-View Spatiotemporal Tracking
- Physics-Aware Gaussians
- Multi-View Feed-Forward Transformer
- Hand-Object Physics-Aware Gaussian-Based Optimization Framework
- Tetrahedral Constraints
- Collision Refinement
- Appearance Decomposition
- VEPHand