好文作者面授招-20141223

  • 主题:Linearized Alternating Direction Method: Two Blocks and Multiple Blocks
  • 主讲人:林宙辰,北京大学
  • 主持人程明明,南开大学
  • 主题:Metric Learning Driven Multi-Task Structured Output Optimization for Robust Keypoint Tracking
  • 主讲人:李玺,浙江大学
  • 主持人白翔,华中科技大学
  • 活动时间:2014年12月23日(周二),北京时间20:30-22:00

1. 相关文献阅读

  • Zhouchen Lin, Risheng Liu, and Zhixun Su, Linearized Alternating Direction Method with Adaptive Penalty for Low Rank Representation, NIPS 2011, arXiv: 1109.0367. (for two block case)
  • Zhouchen Lin, Risheng Liu, and Huan Li, Linearized Alternating Direction Method with Parallel Splitting and Adaptive Penalty for Separable Convex Programs in Machine Learning, accepted by Machine Learning (Special Issue for ACML2013), arXiv: 1310.5035. (for more than two blocks)
  • Metric Learning-Driven Multi-Task Structured Output Optimization for Robust Keypoint Tracking,” Proceedings of Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI), 2015  [pdf] [supp][demo]

2. 讲座资料

2.1 Linearized Alternating Direction Method: Two Blocks and Multiple Blocks. [Slides]

  • 摘要: Alternating direction method (ADM) is an intuitive yet powerful method for various convex optimization problems. However, the traditional ADM assumes that each subproblem is easy to solve and its convergence is proven only in the case of two blocks. Such limitations greatly prevent ADM from wider applications to more complex problems. So I generalize ADM in two aspects. First, I linearize the quadratic penalty term and update the penalty parameter adaptively, introducing linearized ADM (LADM) with adaptive penalty. Second, I modify LADM slightly to account for the multiple block case, introducing linearized ADM with parallel splitting and adaptive penalty. Deeper results are achieved in the scenario of machine learning and signal processing and the proposed algorithms fit for engineering use much better.
  • Bio: ZHOUCHEN LIN (林宙辰) received the Ph.D. degree in applied mathematics from Peking University in 2000. He is currently a Professor at Key Laboratory of Machine Perception (MOE), School of Electronics Engineering and Computer Science, Peking University. He is also a Chair Professor at Northeast Normal University and a guest professor at Beijing Jiaotong University. He was a guest professor at Shanghai Jiaotong University and Southeast University, and a guest researcher at Institute of Computing Technology, Chinese Academy of Sciences. Before joining Peking University, he was a lead researcher at Microsoft Research Asia. He is an associate editor of IEEE T. PAMI and IJCV, an area chair of CVPR 2014, and a Senior member of the IEEE. His webpage is:http://www.cis.pku.edu.cn/faculty/vision/zlin/zlin.htm

2.2 Metric Learning Driven Multi-Task Structured Output Optimization for Robust Keypoint Tracking. [Slides]

  • 摘要: As an important and challenging problem in computer vision and graphics, keypoint-based object tracking is typically formulated in a spatio-temporal statistical learning framework. However, most existing
    keypoint trackers are incapable of effectively modeling and balancing the following three aspects in a simultaneous manner: temporal model coherence across frames, spatial model consistency within frames, and discriminative feature construction. To address this issue, we propose a robust keypoint tracker based on spatio-temporal multi-task structured output optimization driven by discriminative metric learning. Consequently, temporal model coherence is characterized by multi-task structured keypoint model learning over several adjacent frames, while spatial model consistency is modeled by solving a geometric verification based structured learning problem. Discriminative feature construction is enabled by metric learning to ensure the intra-class compactness and inter-class separability. Finally, the above three modules are simultaneously optimized in a joint learning scheme. Experimental results have demonstrated the effectiveness of our tracker.
  • 简介: 李玺:男,博士,浙江大学教授,入选第五批中国国家“青年千人计划”。主要从事计算机视觉、模式识别和机器学习等领域的研究和开发。在目标跟踪、目标行为识别、图像标注、视频检索、哈希(hashing)函数学习等方面取得了深入系统的研究成果,其中在视频的运动跟踪、理解与检索等方面的研究具有特色和优势。 本人在国际权威期刊和国际顶级学术会议发表文章40多篇,包括TPAMI、IJCV、TIP、TKDE、ACM TIST、PR、IVC、ICCV、CVPR、ECCV、ICML、AAAI、ICDM、ACM MM、WWW等。担任神经计算领域知名国际刊物Neurocomputing的Associate Editor,同时担任多个计算机视觉和模式识别方面的国际刊物和国际会议的审稿人和程序委员。获得两项最佳国际会议论文奖(包括ACCV 2010和DICTA 2012),另外分别获得两项中国北京市自然科学技术奖(包括一等奖和二等奖),以及一项中国专利优秀奖。

3. 资料下载

(Visited 213 times, 1 visits today)