Research Interests: Computer vision and graphics
Office Phone: 86-10-6275 7069
Zeng,Gang is an Associate Professor with Tenure at the School of Electronics Engineering and Computer Science in Peking University. He received his B.S. degree from School of Mathematical Sciences at Peking University in 2001, and his Ph.D. degree from Department of Computer Science and Engineering, Hong Kong University of Science and Technology in 2006. He worked as a postdoctoral research fellow in the BIWI Computer Vision Laboratory, ETH Zurich during 2006 to 2008. He joined Peking University in 2008 and was rewarded by "PKU Youth Talent Support Program". His research interests include computer vision and graphics, specifically in image-based scene analysis and modeling.
Dr. Zeng’s works have advanced the state-of-the-arts in image segmentations, object detection, image recognition, shape enhancement and large-scale image search. He has been an associate editor of Neurocomputing since 2013, and he served as area chair of ACPR 2011, CAD/Graphics 2013&2017, and IWRCV 2015. He has published over 20 papers in the top vision and graphics conference and journals, like CVPR, ICCV, IJCV, PAMI, SIGGRAPH, TOG, and his works have been cited over 1800 times based on Google Scholar. He received the Best Poster Award in the 7th Joint Workshop on Machine Perceptions and Robotics 2011, and 4th winner in scene parsing of ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2016.
Dr. Zeng has more than six state research projects with the achievements summarized as follows:
1) Structure-sensitive Image Analysis: In most image analysis tasks the ?rst step towards a successful solution is to discover a robust image representation at a certain scale that helps to map the problem into an objective functional for minimization in a high-dimensional feature space. He has focused on encoding prior structure constraints in the design of image representations at di?erent scales, and made them invariant to inner-class variation and discriminative regarding inter-class di?erence, which boosts the performance of the prior arts.
2) Similarity-aware Shape Enhancement: The classic structure-from-motion (SFM) relies only on sparse feature points to recover scene geometries and camera parameters. The main di?culties in achieving detailed shapes come from the intrinsic ill-posedness of the reconstruction problem, the interference from image noise and insuficient resolution, and the lack of prior shape knowledge. He has been working towards the goal of automatically enhancing shapes with detailed geometries for realistic 3D modeling, by using di?erent shape priors, either learning based or grammar based.
3) Large-scale Data Organization and Clustering: Internet photo/video collections become more and more popular and have been extensively used in modern applications. He also attempted to scale up the conventional data organization and clustering methods for large-scale image analysis and scene modeling applications, e.g., indexing and clustering a large number of SIFT features and other types of visual descriptors, which will promote the utilization of computer vision research findings in large scale real-life applications.