This travel grant supports two US participants to attend the International Workshop on Stochastic Image Grammars (SIG-11). The notion of stochastic image grammars encompasses hierarchical representations of objects and events occurring in images and video, and their associated learning and inference algorithms. The virtue of image grammars lies in their expressive power to represent an exponentially large number of object and event configurations by using a relatively much smaller vocabulary, and a few compositional rules.
Statistics, machine learning, natural language processing, and cognitive psychology experience a resurgence of stochastic grammars. In computer vision, however, this momentum seems to be present only in the area of 2D object recognition. The main objective of the workshop is to promote interdisciplinary research among these traditionally separate scientific disciplines toward grammar-based formulations of a wider range of vision problems, beyond object recognition, such as, e.g., 3D structure from motion, and activity recognition. The workshop is also aimed at reducing the apparent disconnect between research groups working on image grammars, by addressing the need for a unified theoretical framework. To this end, SIG-11 provides a forum for sharing research experiences in grammars between the vision community and the keynote speakers who are experts in cognitive psychology, neuroscience, and natural language processing. Solicited peer-reviewed papers are expected to be published in the proceedings of the 13th International Conference on Computer Vision.
This project was aimed at organizing a workshop on Stochastic Image Grammars in conjunction with 13th International Conference of Computer Vision (ICCV), in Barcelona, Spain, 2012. The notion of stochastic image grammars encompasses hierarchical representations of objects and events, semantic and spatiotemporal contexts, taxonomy of visual categories, and their associated learning and inference algorithms. The project activities included the following: organization of the workshop, preparation of the workshop program, and making a webpage for broader dissemination of the workshops ideas and results. The main objective of the workshop was to promote interdisciplinary research in statistics, applied mathematics, natural language processing, artificial intelligence, and computer vision. The workshop was organized for one full day, and included oral and poster presentations of the peer reviewed papers, talks of invited speakers, and the workshop's panel discussion. The workshop hosted the following renowned invited speakers: - Prof. Pedro Domingos from University of Washington, U.S.A. - Prof. Roxana Girju from University of Illinois Urbana Champaign, U.S.A. - Prof. Rama Chellappa from University of Maryland, U.S.A. - Prof. Pedro Felzenszwalb from Brown University, U.S.A. - Prof. Edwin Hancock from University of York, U.K. - Prof. Noah Goodman from Stanford University, U.S.A. Findings: The workshop's papers and invited talks have advanced the state of the art in terms of introducing new: - Formulations of temporal stochastic image grammars for representation and recognition of human activities occurring in videos - Inference and learning algorithms for context-sensitive stochastic image grammars - Formulations for fusing text and video data using stochastic image grammars The workshop's papers and abstracts of invited talks were published in the Proceedings of ICCV 2011, editors: Sinisa Todorovic, Song-Chun Zhu, and Ales Leonardis Bibliography: http://ieeexplore.ieee.org/xpl/most RecentIssue.jsp? The workshop's URL is: http://vcla.stat.ucla.edu/sig11/home.html The website allows easy access to the objectives, program, peer-reviewed papers, and invited talks of the workshop. Contributions within Discipline: - The workshop extended the previous work on stochastic image grammars on 2D object recognition to new vision problems, including 3D structure from motion, and activity recognition. - The workshop also established theoretically sound connections between grammar representations and probabilistic first-order logic. Typically, for the purposes of inference, existing work transforms predicates of a probabilistic first- order logic into augmented formulas in the conjunctive normal form (CNF). These AND-OR formulas essentially represent grammar rules, augmented with existential and universal quantifiers. The relations between first-order logic predicates and CNF formulas are poorly understood, and the workshop's peer-reviewed papers provided new insights into this problem. Contributions to Other Disciplines: One of the workshop's topics was integration of text and video information for robust intelligence. The workshop will have an immediate impact on the natural language processing community, represented by the invited speaker Prof. Roxana Girju from Univ. of Illinois Urbana Champaign.
