The objective of this work is to analyze the behavior of deep neural networks that use single images to solve two different computer vision tasks: depth estimation and 3D reconstruction, both traditionally solved using stereo images and epipolar geometry, or structure from motion.

​

Several systems have accomplished these tasks using single images with considerable accuracy. However, these systems do not discuss how the neural networks achieve these results and what they learn. This work is based on two papers: "How do neural networks see depth in single images?" by Dijk and Croon, and "What Do Single-view 3D Reconstruction Networks Learn?" by Tatarchenko et al.

​

The presentation is available in the following link:

​

​