Reconstructing Animatable 3D Categories from Videos

[Back]

Ablation on Morphology Code Regularization

Without morphology code regularization, the reconstructed shape may look reasonable from the reference viewpoint, but severely distorted from a novel viewpoint. We posit the body parts that are not well-covered in the video are inherently difficult to infer. As a result, the shapes become degenerate without relying on priors from other videos in the dataset.
Ours
without β code regularization