Large-Scale High-Quality 3D Gaussian Head Reconstruction from Multiview Captures
6 hours ago
- #Multi-View
- #3D Reconstruction
- #Gaussian Heads
- HeadsUp is a scalable feed-forward method for reconstructing high-quality 3D Gaussian heads from multi-camera setups.
- It uses an encoder-decoder architecture to compress input views into a latent representation, decoded into UV-parameterized 3D Gaussians anchored to a neutral head template.
- The UV representation decouples 3D Gaussian count from input image resolution, enabling training with many high-resolution views.
- Trained on an internal dataset of over 10,000 subjects, it achieves state-of-the-art quality and generalizes to novel identities without test-time optimization.
- The method supports downstream applications like generating novel 3D identities and animating heads with expression blendshapes.
- It is evaluated on a large-scale multi-camera capture dataset and Ava-256, demonstrating riggable and animatable reconstructions.