Tue, Feb 28, 2023 @ 03:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
Title: Learning to Optimize the Geometry and Appearance from Images.
The ability to infer geometry and appearance from images impacts various applications such as AR/VR, autonomous driving, and more. Compared to traditional methods, deep convolutional neural networks have proven to be more robust and accurate. However, the practical use of deep learning in these applications still faces three major challenges: (1) the acquisition of 3D training data; (2) the development of a fast, robust, and accurate 3D vision framework; (3) the integration of complex 3D representations into the neural network.
To address these challenges, my research focuses on optimization techniques in the context of deep learning. Specifically, when paired 2D and 3D data is not available, we propose a differentiable rendering framework that allows neural networks to learn 3D shapes directly from 2D images. On the other hand, when full supervision is available, we develop a framework that trains a neural network to optimize the target representation and demonstrate the performance on the vanishing point detection task. Finally, we explore the face avatar creation task and propose dense visual-semantic correlation on top of a semantically-aligned UV space to effectively integrate complex 3D representations into the neural optimization framework. Our neural optimization techniques help to develop practical 3D computer vision systems.
Committee members are Randall Hill, Andrew Nealen, Aiichiro Nakano, Stefanos Nikolaidis, and Yajie Zhao.
Audiences: Everyone Is Invited
Contact: Asiroh Cham
Event Link: https://usc.zoom.us/j/3154287574