报告时间:2024年12月16日上午10:20
报告地点:浙江大学紫金港校区蒙民伟楼402会议室
报告题目: Unconventional Visual Information-Driven Deep Learning Practices
报告人:白潇洋 博士
主持人:崔兆鹏 研究员
Abstract: In contrast to RGB images, unconventional visual information such as polarization images, event streams, depth maps and CT/MRI images are less explored by the CV/CG community and they are often processed naively with neural network models that are designed to learn from intensity images captured by ordinary cameras. Such practice neglects the unique optical and physical properties of those input modalities and may result in suboptimal performance, reduced generalizability and inapplicability to real-world scenarios. Therefore, it is of crucial importance to develop deep learning methods that are aware of the properties of unconventional visual modalities. In this talk, I will briefly introduce three representative researches in this direction: underwater localization based on omnidirectional polarization images, RGBD imaging via asymmetrically focused stereo cameras, and dynamic 3D Gaussian splatting empowered by event streams.
Bio: Xiaoyang Bai is currently a postdoctoral fellow at the University of Hong Kong, working with Prof. Evan Peng and Prof. Edmund Lam. He received his Ph.D. degree from University of Illinois Urbana-Champaign (UIUC) and his bachelor degree from University of California, Berkeley (UCB). His research centers around polarization-based vision, event-based vision, 3D scene reconstruction and underwater imaging. He serves as a reviewer for multiple international journals and conferences including ACM TOG and IEEE VR.