CVPR 2010 Tutorial on Higher Order Models in Computer Vision
Speakers Carsten Rother (MSR Cambridge) Sebastian Nowozin (MSR Cambridge) (Note, unfortunately Pushmeet Kohli cannot come to CVPR) Slides online now All slides in pdf format: download or in powerpoint format: download You are welcome to use these slides for talks. Please give appropriate credits.
Purpose of this course Many labelling problems in computer vision such as image restoration, disparity estimation and object recognition are often modelled via Markov Random Fields. Most commonly the model has an underlying simple 4(8)-connectivity field. These simple models are very popular, very likely due to the fact that efficient inference (and learning) techniques exist. It is well known that modelling jointly several variables, i.e. higher-order cliques, considerably improve the modelling power and hence the results. The goal of this tutorial is to analyse and categorize various types of different higher-order random field models which have been considered in the past (e.g. patch-based priors, curvature prior, topology prior, etc). The key question for such powerful models is whether efficient and powerful inference techniques exist. This question is the main focus of the tutorial, and we will review recent work which has shown that inference is indeed tractable, by e.g. transforming the higher-order model into a pair-wise one. Relationship to tutorial at ICCV 2009 We had given a general (full day) tutorial at ICCV 09 on MAP Inference in Discrete Models. The ICCV tutorial was very general and higher-order models were not really covered. Given recent advances and interest in this field we believe that this tutorial will appeal to a large audience. We hope to inspire many people to use in the future more sophisticated higher-order models. Final Syllabus 1. Introduction to Higher Order Random Field Models 2. Background of inference techniques: Pseudo-Boolean Optimization, Message Passing, LP-relaxation techniques, and Dual Decomposition. 3. Low to medium order Models a. Patch based models: FoE, Pattern-based potentials. b. Region-based potentials (label-consistency; P^n Potts) c. Curvature, etc. d. Transformation of a+b to pair-wise model. e. Examples: stereo, denoising, object recognition, etc. 4. Global (full image) Models a. Topology: Connectivity, Bounding Box, Silhouette constraint etc. b. Preserve distribution of labels (Marginal Probability Field), and global appearance models. c. Label cost prior. d. Examples: de-noising, interactive segmentation, co-segmentation, FilterFlow etc. 5. Summary and Directions for Future Work About the Speakers
		Carsten Rother received his Diploma degree with distinction in 1999 at the University of Karlsruhe/Germany. He did his PhD at the Royal Institute of Technology Stockholm/Sweden, supervised by Stefan Carlsson and Jan-Olof Eklundh. Since 2003 he is a researcher at Microsoft Research Cambridge/UK. He supervises several PhD students and gives frequently invited talks, organizes workshops and taught a tutorial (MAP Inference in Discrete Models at ICCV 09). His research interests are in the field of “Markov Random Fields for Computer Vision”, “Discrete Optimization”, and “Vision for Graphics”. He has published more than 20 high impact papers (at least 10 citations) at international conferences and journals. He won the best paper honourable mention award at CVPR ’05, and he was awarded the DAGM Olympus price 2009. He serves on the program committee of major conferences (e.g. SIGGRAPH, ICCV, ECCV, CVPR, NIPS), and has been area chair for BMVC ’08 – ‘10 and DAGM ‘10.
		Sebastian Nowozin is a researcher in the Machine Learning and Perception group at Microsoft Research Cambridge. He received his Master of Engineering degree from the Shanghai Jiaotong University and his diploma degree in computer science with distinction from the Technical University of Berlin in 2006. He received his PhD degree with highest distinction in 2009 for his thesis on learning with structured data in computer vision, completed at the Max Planck Institute for Biological Cybernetics, Tuebingen and the Technical University of Berlin. His research interest is diverse and includes computer vision, machine learning, and continuous and discrete optimization. He organizes the successful “Optimization for Machine Learning” workshop series at NIPS (OPT 2008, OPT 2009) and serves as PC-member/reviewer for machine learning (e.g. NIPS, ICML, AISTATS, UAI, ECML, JMLR) and computer vision (e.g. CVPR, ICCV, ECCV, PAMI, IJCV) conferences/journals.