Skip to content Skip to navigation
Computer Science Department Colloquium
9/23/2014 11:00 am
CoRE Lecture Hall (Room 101)

Composite Statistical Inference in Semantic/Video Segmentation

Fuxin Li, Georgia Institute of Technology

Faculty Host: Chao Chen and Dimitris Metaxas

Abstract

I will present a summary of the research that I have done with my collaborators in the past few years on semantic and video segmentation. Different from conventional localized CRF approaches, we propose to utilize segment proposals generated from unsupervised image segmentation, and focus on predicting/utilizing statistics on these segments for learning and inference. Focusing on subset statistics avoids the need to consider normalization constants in both learning and inference, and learning becomes a conventional regression problem. Detailed learning/inference procedures will then be presented on the semantic segmentation and unsupervised video segmentation problems. During the inference, continuous parameters are defined on superpixels obtained by multiple intersections of segments, then the optimal segments are outputted from the inferred superpixel statistics. The algorithms are capable of recombine and refine initial mid-level proposals, as well as handle multiple interacting objects. In the PASCAL VOC segmentation challenge, the proposed approach obtains high accuracy and successfully handles images of complex object interactions. In video segmentation it improves the state-of-the-art by a large margin.

Bio

Dr. Li is interested in the intersection of machine learning and computer vision, espeically segmentation in images/videos  and visual object recognition. He has published more than 30 papers in machine learning, computer vision and proteomics. He and his colleagues have won the prestigious PASCAL Visual Object Recognition challenge in Segmentation from 2009-2012, beating teams from University of Stanford, University of Chicago, Oxford University, University of California—Berkeley, among others. He is a reviewer for more than 15 renowned international journals, and serves as the program committee for all major machine learning and computer vision conferences such as CVPR, ICCV, ECCV, NIPS, ICML. He has won a Microsoft research award, 2 best reviewer awards and is currently leading an NSF project.