20161012-34 周晓巍：3D Object Geometry from Single Image

报告嘉宾1: 周晓巍  （宾夕法尼亚大学（University of Pennsylvania））

报告时间：2016年10月12日（星期三）晚20：00（北京时间）

报告题目：3D Object Geometry from Single Image

主持人：张姗姗（德国马普所）

报告摘要：

The past few years have witnessed remarkable advancements in 2D image understanding driven by deep learning which enable us to reliably classify an image, detect objects in it, semantically label the pixels and even automatically generate captions. However, 2D information alone is not sufficient when it comes to any scenario involving interactions between human, robot and the world, where 3D geometric properties about the scene is necessary. Given the expenses of depth sensors and their limitations in certain scenarios, recovering the 3D geometry from RGB images has drawn a lot of attention. This talk focuses on our recent progresses on recovering 3D object geometry including pose and structure of both rigid and articulated objects from a single image, levering data-driven representations learned from both 2D and 3D data. Specifically, I will first show that, using semantic keypoints and object CAD models, it is feasible to accurately and efficiently estimate the 6-DoF pose of a textureless object with a high precision that allows a robot to grasp the object. In the second part, I will propose an EM algorithm that combines a sparse representation with a CNN to reconstruct the 3D human pose from a monocular image or video and accounts for the uncertainty in the CNN output. Finally, I will introduce how to reconstruct an object-class model from images of different instances with a multi-image matching algorithm that optimizes the global matching consistency.

参考文献：

Sparse Representation for 3D Shape Estimation: A Convex Relaxation Approach.

X. Zhou, M. Zhu, S. Leonardos, K. Daniilidis.

IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 2016.

Sparseness Meets Deepness: 3D Human Pose Estimation from Monocular Video.

X. Zhou, M. Zhu, S. Leonardos, K. Derpanis, K. Daniilidis.

IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.

Multi-Image Matching via Fast Alternating Minimization.

X. Zhou, M. Zhu, K. Daniilidis.

International Conference on Computer Vision (ICCV), 2015.

Single Image Pop-Up from Discriminatively Learned Parts.

M. Zhu, X. Zhou, K. Daniilidis.

International Conference on Computer Vision (ICCV), 2015.

报告人简介:

Xiaowei Zhou is a Postdoctoral Researcher in the Computer and Information Science Department at University of Pennsylvania. His research interests are on 3D vision problems including object pose estimation, shape reconstruction, human pose estimation and image matching. His current work attempts to combine 3D geometry, optimization and learning methods to extract semantic and geometric information of 3D environment from visual data. He obtained his Bachelor's degree in Optical Engineering from Zhejiang University, 2008, and his PhD degree in Electronic and Computer Engineering from The Hong Kong University of Science and Technology, 2013.