Unsupervised Learning of Probably Symmetric Deformable 3D Objects From Images in the Wild (Invited Paper)

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Learning 3D object properties from 2D images is challenging, especially without supervision.
• Existing methods often require extensive labeled data or 2D correspondences.
• Exploiting inherent object properties like symmetry can aid unsupervised learning.

Purpose of the Study:

• To develop an unsupervised method for learning 3D deformable object categories from single images.
• To disentangle image factors including depth, albedo, viewpoint, and illumination.
• To leverage object symmetry for robust 3D shape recovery.

Main Methods:

• An autoencoder architecture is employed to factorize input images.
• Unsupervised disentanglement is achieved by exploiting approximate object symmetry.
• Illumination reasoning and a learned symmetry probability map handle non-symmetric appearances and probabilistic symmetry.

Main Results:

• The method accurately recovers 3D shapes of human faces, cat faces, and cars from single-view images.
• No external supervision or prior shape models are required.
• Achieves superior accuracy compared to supervised methods using 2D correspondences on benchmarks.

Conclusions:

• Unsupervised learning of 3D deformable object categories is feasible by leveraging symmetry.
• The proposed method effectively disentangles intrinsic image properties for shape reconstruction.
• This approach offers a promising direction for 3D understanding from limited visual data.