UC Riverside

Visual analysis is a fundamental task to develop approaches to understand contents. With the advances of deep learning models, visual image synthesis plays an increasingly important role because it leverages generative models for synthesizing novel images which can be used for training and testing. This dissertation presents new techniques for visual analysis and synthesis by developing novel deep learning techniques, in particular methods for sports analytics, vision and language, face synthesis, scene perceptions, human body mesh understanding and visual interpretations of generative models. This dissertation deals with sports analytics, image and language pre-training, translation of facial attributes, passive depth estimation in the indoor environments, human body mesh reconstruction and visual interpretations of variational autoencoders. It develops a range of advanced techniques for dribbling and goal recognition, language-supervised contrastive learning for visual understanding, translating features on a human face, passive range application of AR/VR, reconstructing human body mesh and interpreting variational autoencoders. Both theory and experimentation will be presented.