Reconstructing 3d Humans From Images

The past decade we have seen remarkable progress in Computer Vision, fueled by the recent advances in Deep Learning. Unsurprisingly, human perception has been the center of attention. We now have access to systems that can work remarkably well for traditional 2D tasks like segmentation or pose estimation. However, scaling this to 3D remains particularly challenging because of the inherent ambiguities and the scarcity of annotations. The goal of this dissertation is to describe our contributions towards automating 3D human reconstruction from images. First, we will explore the use of different representations for human mesh recovery, discuss their advantages and show how they can be useful for learning deformations beyond standard parametric body models. Next, motivated by the limited availability of annotated data, we will present a method that leverages a collaboration between regression and optimization methods to successfully address this. Subsequently, we will describe our work on modeling the ambiguities in 3D human reconstruction and demonstrate its usefulness for solving a variety of downstream tasks, such as human body model fitting. Last, we will move beyond single-person 3D pose estimation and show how we can scale our methods to work on challenging real-world scenes with multiple humans