智能感知与计算系列讲座
Lecture Series in Intelligent Perception and Computing
题 目(TITLE):Learning Disentangled Representations for 3D View Synthesis and Shape Reconstruction
讲 座 人(SPEAKER): Dr. Jimei Yang
主 持 人 (CHAIR):Dr. Ran He
时 间 (TIME):Jan 11, 2017 (Wendsday), 15:00-16:00
地 点 (VENUE):Meeting Room (1610), 16 Floor, Intelligent Building
报告摘要(ABSTRACT):
3D view synthesis and shape reconstruction are important problems in both vision and graphics. In this talk, we present action-conditioned recurrent convolutional networks to replicate the sequential 3D rotations of an object from a single image. The recurrent structure allows the network to capture long-term dependencies of viewpoint changes that in turn leads to the disentangling of latent factors (identity and pose) without direct supervision. We demonstrate the quality of view synthesis on natural human faces and rendered 3d shapes, and as well the performance of learned identity features for fine-grained object recognition, with comparisons to supervised convolutional networks. Furthermore, based on the learned view-invariant identity representations, we develop Perspective Transformer Networks to reconstruct the 3D shape from a single view of an object. The in-network implementation of perspective camera geometry introduces a novel projection loss for training the PTNs without 3D supervision.
报告人简介(BIOGRAPHY):
Jimei Yang is a research scientist at Adobe Research. He received his Ph.D. degree from the University of California, Merced in 2015, supervised by Prof. Ming-Hsuan Yang. He was a visiting graduate student in Prof. Honglak Lee’s group at the University of Michigan, Ann Arbor. He obtained his Master’s degree from USTC. His research focuses on structured deep learning with applications in vision and graphics.
|