Real-time 3D Hand Reconstruction in Challenging Scenes from a Single Color or Depth Camera

Abstract

Hands are one of the main enabling factors for performing complex tasks and humans naturally use them for interactions with their environment. Reconstruction and digitization of 3D hand motion opens up many possibilities for important applications. Hands gestures can be directly used for human–computer interaction, which is especially relevant for controlling augmented or virtual reality (AR/VR) devices where immersion is of utmost importance. In addition, 3D hand motion capture is a precondition for automatic sign-language translation, activity recognition, or teaching robots. Different approaches for 3D hand motion capture have been actively researched in the past. While being accurate, gloves and markers are intrusive and uncomfortable to wear. Hence, markerless hand reconstruction based on cameras is desirable. Multi-camera setups provide rich input, however, they are hard to calibrate and lack the flexibility for mobile use cases. Thus, the majority of more recent methods uses a single color or depth camera which, however, makes the problem harder due to more ambiguities in the input. For interaction purposes, users need continuous control and immediate feedback. This means the algorithms have to run in real time and be robust in uncontrolled scenes. These requirements, achieving 3D hand reconstruction in real time from a single camera in general scenes, make the problem significantly more challenging. While recent research has shown promising results, current state-of-the-art methods still have strong limitations. Most approaches only track the motion of a single hand in isolation and do not take background-clutter or interactions with arbitrary objects or the other hand into account. The few methods that can handle more general and natural scenarios run far from real time or use complex multi-camera setups. Such requirements make existing methods unusable for many aforementioned applications. This thesis pushes the state of the art for real-time 3D hand tracking and reconstruction in general scenes from a single RGB or depth camera. The presented approaches explore novel combinations of generative hand models, which have been used successfully in the computer vision and graphics community for decades, and powerful cutting-edge machine learning techniques, which have recently emerged with the advent of deep learning. In particular, this thesis proposes a novel method for hand tracking in the presence of strong occlusions and clutter, the first method for full global 3D hand tracking from in-the-wild RGB video, and a method for simultaneous pose and dense shape reconstruction of two interacting hands that, for the first time, combines a set of desirable properties previously unseen in the literature.