Probably symmetric: 3D objects learned from raw 2D images

What it does This is the code behind a CVPR 2020 Best Paper. Feed it a pile of single-view images—celebrity faces, cat portraits, or car photos—and it learns to reconstruct any new 2D snapshot as a textured 3D mesh plus camera viewpoint, all without depth maps, landmark labels, or multi-view stereo. The repository ships with pretrained checkpoints and training configs for human faces, synthetic faces, cat faces, and synthetic cars.

The interesting bit The method bets on the idea that most natural objects are probably symmetric, using that weak symmetry prior to disentangle shape, albedo, illumination, and pose from a single flat image. It is also a rare best-paper release whose demo script will grudgingly run on a CPU, even if its neural rendering backend much prefers a GPU.

Key highlights

• Winner of the CVPR 2020 Best Paper Award.
• Learns deformable 3D categories from raw single-view images with zero 3D supervision.
• Requires no multi-view data, 2D/3D keypoints, or prior shape models.
• Includes pretrained models for faces, cats, and cars.
• Demo inference can run without a GPU; training and full testing require one.

Caveats

• The dependency stack is frozen in time: PyTorch 1.2.0 and CUDA 9.2, so modern environments need careful reconstruction.
• The neural_renderer package is GPU-only and can fail to compile on GCC versions below 5.
• Automatic face detection in the demo is limited to human faces; cats and cars must be cropped manually.

Verdict Researchers exploring unsupervised single-view 3D reconstruction will find a clean reference implementation here. Everyone else should treat it as a reproducibility artifact rather than a drop-in library.