Abstract:
Hands are the central means by which humans interact with their surroundings. Understanding human hands helps us analyze human behavior, perform other visual analysis tasks such as action and gesture recognition, and develop applications for metaverse and robotics. I study several problems in this thesis to understand human hands in visual data. The first step toward the visual understanding of human hands is to detect hands, and I propose a contextual attention method to localize hands in images. While localizing hands is essential, more than detection is needed to understand hand semantics. To better understand hand interactions, we need to understand hand contact information, and I address this in my second work. For a scene containing multiple people, in addition to localizing hands and recognizing their contact, we need to understand what object is manipulated by whom and which person performs what activity. This requires us to analyze human hands and body jointly, and I address this hand-body association task in my third work. Human activities are transient, allowing hands to move across time. In my fourth work, I propose an online tracking method to track multiple hands in videos to understand how hands move over time. When hands move, they contact, interact with, hold, and carry objects. Therefore, identifying, localizing, and tracking hand-held objects is crucial, and I address this in my fifth work. Generative Artificial Intelligence (GenAI) has several applications in content creation, and one of the most popular ways to create content is using text-to-image methods. While text-to-image methods can synthesize high-quality humans, they often struggle to generate realistic hands. In my final work, I propose a method to generate images with realistic hands using hand priors.