Unsupervised Learning of Probably Symmetric Deformable 3D Objects from Images in the Wild

---

The researchers assumed that many objects have a symmetric structure. Based on that assumption, 3d object that contains depth, albedo, viewpoint and illumination of a single image can be achieved without supervision. → using symmetry, they used the differences of appearance of two sides of a single object, to learn those features.

Introduction

---

The research was done with two underlying conditions:

• No 2D or 3D ground truth information is available: removes the bottleneck of collecting image annotations.
• Algorithm must use an unconstrained collection of single-view images: can use various images of a deformable object category.

This was formulated by autoencoder.