Deep Learning for Analysis and Reconstruction of 3D Shapes as Point Clouds and Meshes

The world exists in three dimensions (3D) and it is how we humans perceive it. Our superiority in processing 3D visual information allows us to understand the underlying 3D geometry, and then respond to events around us effectively. In the last decade, fueled by deep learning and deep neural networks, intelligent systems have revolutionized many areas in computer vision, including 3D vision. In this talk, I will focus on 3D reconstruction from 2D images using deep learning-based solutions. First, I will introduce a 3D reconstruction system that can predict precise 3D point clouds given only a single image. The solution deforms a randomly initialized point cloud into the object shape described in the 2D image using a new point cloud operator. Then, I will discuss a method that reconstructs human 3D avatars with a high level of detailed garments from multi-view images. The method is inspired by the smoothing process of surfaces, and uses a new graph diffusion operator to reverse the smoothing process.