Moderator: Boxin Shi, Peking University, China
Boxin Shi received the BE degree from the Beijing University of Posts and Telecommunications (BUPT), the ME degree from Peking University (PKU), and the PhD degree from the University of Tokyo, in 2007, 2010, and 2013. He is currently a Boya Young Fellow Assistant Professor and Research Professor at PKU, where he leads the Camera Intelligence Group. Before joining PKU, he did postdoctoral research with MIT Media Lab, Singapore University of Technology and Design, Nanyang Technological University from 2013 to 2016, and worked as a researcher in the National Institute of Advanced Industrial Science and Technology from 2016 to 2017.

Lecturer(s): Boxin Shi, Peking University, China
Boxin Shi received the BE degree from the Beijing University of Posts and Telecommunications (BUPT), the ME degree from Peking University (PKU), and the PhD degree from the University of Tokyo, in 2007, 2010, and 2013. He is currently a Boya Young Fellow Assistant Professor and Research Professor at PKU, where he leads the Camera Intelligence Group. Before joining PKU, he did postdoctoral research with MIT Media Lab, Singapore University of Technology and Design, Nanyang Technological University from 2013 to 2016, and worked as a researcher in the National Institute of Advanced Industrial Science and Technology from 2016 to 2017.

Description: Modern 3D computer vision methods, represented by multi-view stereo and structure-from-motion, have achieved faithful 3D reconstruction from a set of images. But are the reconstruction quality and density really sufficient for your purpose? Despite requiring more controlled setups than multi-view stereo, photometric approaches have proven to be invaluable tools in applications such as Hollywood movies, industrial quality inspection, etc. since they can reconstruct fine surface details at superior quality. This course covers a thread of ‘photometric’ approaches to high-fidelity 3D reconstruction, which enable truly dense 3D estimation at the level of pixel-level details from shading observations. The course will mainly cover photometric stereo techniques that take as input a set of images observed under different illumination conditions from a fixed viewpoint to compute the shape in the form of surface normals with the same high resolution as the 2D image. While conventional photometric stereo methods make various assumptions over reflectance and illumination, they are being relaxed in modern methods by powerful machine learning approaches so as to be practical in diverse scenarios. In addition, newly rendered datasets and captured real world datasets have been proposed for training and testing data-driven approaches for photometric stereo, which shows superior performance over non-learning approaches. This course will also cover basic principles about how conventional and non-learning based photometric stereo works for a self-contained purpose.