國立清華大學資訊工程學系

Department of Computer Science

National Tsing Hua University

專題演講

SEMINAR

主講人：Wei-Lun (Harry) Chao (趙偉崙)

SPEAKER (USC)

題　目：Zero-shot Learning for Visual Recognition in the Wild

TOPIC  

時  間： 106年01月05日(四) 11:00AM - 12:00PM

DATE　　

地　點： 台達館102室

PLACE

                                                                                                                       聯絡人：陳煥宗教授

Abstract
The availability of large-scale labeled images is one key factor that contributes to recent successes in visual object recognition. It is well-known, however, that object frequencies in natural images follow long-tailed distributions: most of the objects “in the wild” do not occur frequently enough for us to collect and label a large set of representative images for building high-quality classifiers. Zero-shot learning (ZSL), which aims to expand the classifiers and labeling space from seen objects to unseen objects, has since emerged as a promising paradigm to remedy the above difficulty.
In this talk, I will present our recent work in ZSL, first on how to effectively leverage the class semantic descriptions in relating seen and unseen objects. Viewing object classes as coordinates in both the semantic and visual model space, we develop algorithms to synthesize classifiers and predict visual exemplars for unseen classes, achieving promising results on ImageNet with over 20K unseen classes. Then I will take a step away from algorithms to the problem setting of ZSL. Specifically, we investigate generalized ZSL (GZSL), which relaxes the unrealistic assumption in conventional ZSL that excludes seen objects in the test phase. We introduce an effective calibration method and develop an evaluation metric to characterize the trade-off in recognizing seen and unseen objects. We further establish a performance upper bound on GZSL, suggesting that improving semantic descriptions is vital for zero-shot learning.

Bio
Wei-Lun (Harry) Chao is a 4th-year Ph.D. student in Computer Science at University of Southern California, where he is advised by Professor Fei Sha. He received a M.S. and a B.S. in Communication Engineering at National Taiwan University and National Chiao Tung University, respectively. His research interests are in computer vision and machine learning. He is particularly interested in transfer learning, structure learning, and semantic modeling. His personal website is http://www-scf.usc.edu/~weilunc/index.html