Reproduce “Learning the 3D Fauna of the Web” 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
Zizhang Li, Dor Litvak, Ruining Li, Yunzhi Zhang, Tomas Jakab, Christian Rupprecht, Shangzhe Wu, Andrea Vedaldi and 1 more
3D-Fauna learns a quadruped model from Internet images and predicts a posed 3D mesh from one image. It uses self-supervised image features to combine base shapes in a Semantic Bank without species labels. On APT-36K horse keypoint evaluation, it scores [email protected] of 0.853, versus 0.402 without the bank. This 2D keypoint result does not establish accurate 3D geometry across species.
Key insights
- The Semantic Bank of Skinned Models (SBSM) is critical for pan-category generalization, enabling the model to learn and interpolate shapes for over 100 diverse species without explicit category labels.
- A mask discriminator, conditioned on semantic features and applied to randomly rendered views, effectively mitigates photographer bias in 2D internet data, preventing the generation of elongated or unrealistic 3D shapes.
- Self-supervised image features provide sufficient semantic correspondence to learn articulated, deformable 3D models from uncurated 2D internet images across a wide range of animal categories, even for species with limited data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.