Towards accurate multi-person pose estimation in the wild

Abstract

In this thesis we are concerned with the problem of articulated human pose estimation and pose tracking in images and video sequences. Human pose estimation is a task of localising major joints of a human skeleton in natural images and is one of the most important visual recognition tasks in the scenes containing humans with numerous applications in robotics, virtual and augmented reality, gaming and healthcare among others. Articulated human pose tracking requires tracking multiple persons in the video sequence while simultaneously estimating full body poses. This task is important for analysing surveillance footage, activity recognition, sports analytics, etc. Most of the prior work focused on the pose estimation of single pre-localised humans whereas here we address a case with multiple people in real world images which entails several challenges such as person-person overlaps in highly crowded scenes, unknown number of people or people entering and leaving video sequences. The first contribution is a multi-person pose estimation algorithm based on the bottom-up detection-by-grouping paradigm. Unlike the widespread top-down approaches our method detects body joints and pairwise relations between them in a single forward pass of a convolutional neural network. Multi-person parsing is performed by optimizing a joint objective based on a multicut graph partitioning framework. Secondly, we extend our pose estimation approach to articulated multi-person pose tracking in videos. Our approach performs multi-target tracking and pose estimation in a holistic manner by optimising a single objective. We further simplify and refine the formulation which allows us to reach close to the real-time performance. Thirdly, we propose a large scale dataset and a benchmark for articulated multi-person tracking. It is the first dataset of video sequences comprising complex multi-person scenes and fully annotated tracks with 2D keypoints. Our fourth contribution is a method for estimating 3D body pose using on-body wearable cameras. Our approach uses a pair of downward facing, head-mounted cameras and captures an entire body. This egocentric approach is free of limitations of traditional setups with external cameras and can estimate body poses in very crowded environments. Our final contribution goes beyond human pose estimation and is in the field of deep learning of 3D object shapes. In particular, we address the case of reconstructing 3D objects from weak supervision. Our approach represents objects as 3D point clouds and is able to learn them with 2D supervision only and without requiring camera pose information at training time. We design a differentiable renderer of point clouds as well as a novel loss formulation for dealing with camera pose ambiguity.