• 学院公告

刘念军博士学术讲座预告
作者:null 发布日期:2010-12-22 浏览次数:

 

报告题目:

高效监督分割方法与高阶图割立体匹配理论

(Efficient Supervised Segmentation and Higher-order Graph Cut Stereo Matching)

 

报告人:  

刘念军博士 (澳大利亚国家信息与通信技术研究中心研究员/澳洲国立大学博导)


报告时间:  2010年12月22日(星期三),下午2:30


报告地点: 郁B307

 

欢迎本院有兴趣的广大师生参加!

 

报告人简介:

Dr. Nianjun Liu received his ME degree from University of Science and Technology of China (USTC) in 1999, and was awarded his Ph.d degree in 2004 and postdoctoral in 2005 from Department of Information Technology and Electrical Engineering  at  University of  Queensland, Australia.  Since 2005 till now, he has been working as a research scientist in National Information and Communications Technology Institute of Australia (NICTA). Dr. Liu is also an adjunct research fellow and the supervisor of Ph.d research students at College of Engineering and Computer Science of The Australian National University. In the past five years, Dr. Liu has done up to 40 peer-reviewed publications. He is a key member of visual processing division of Australian Bionic eye initiative which is a national top innovative project for summit 2020 Australia. (http://www.bionicvision.org.au/)

His research interests including computer vision and machine learning, focusing on segmentation from supervised learning and superpixel-based higher order stereo matching. He is a member of IEEE and ACM society.

 

报告摘要:
The presentation covers two sub-topics.

1) Efficient segmentation learned from structural support vector machine: Conditional random field methods (CRFs) have gained popularity for image labeling tasks in recent years, we describe an alternative discriminative approach, by extending the large margin principle to incorporate spatial correlations among neighboring pixels. In particular, by explicitly enforcing the submodular condition, graph cuts is conveniently integrated as the inference engine to attain the optimal label assignment efficiently. In addition to node and edge feature functions for enforcing local label consistency, our algorithm is shown to be capable of readily incorporating higher-order scene context. Empirical studies on a variety of image datasets suggests that our approach performs competitively comparing to the state-of-the-art scene labeling methods. 

2) Segment-based stereo matching by higher order graph cuts: upon on state-of-the arts over-segment methods, an efficient approach is developed for dense stereo matching through robust higher-order MRFs and graph cut based optimization, which combines the conventional data and smoothness terms with the robust higher-order potential term. The experimental results on real-scene data sets clearly demonstrate that our over-segment-based higher order stereo matching approach outperforms conventional stereo matching algorithms, as well as how over-segments improve the stereo matching process.