Reproduce “NoPo-Avatar: Generalizable and Animatable Avatars from Sparse Inputs without Human Poses” 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
Jing Wen, Alexander G. Schwing, Shenlong Wang
NoPo-Avatar introduces a method to reconstruct generalizable and animatable 3D human avatars from sparse input images without requiring any human or camera pose information during test-time. This approach achieves high visual fidelity and robust performance, outperforming pose-dependent baselines, particularly in real-world scenarios with noisy pose estimates.
Key insights
- Decoupling avatar reconstruction from explicit pose guidance at test-time significantly enhances robustness and practical applicability in real-world settings with noisy pose data.
- The dual-branch design effectively combines global structural coherence from the template branch with fine-grained texture capture from the image branches, enabling comprehensive avatar reconstruction and inpainting capabilities.
- Specific auxiliary losses, particularly the projection loss (L_proj) for pixel-aligned Gaussians and the LBS loss (L_lbs) for correct canonical-pose reconstruction, are critical for achieving high-fidelity and animatable results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.