VASC Seminar: Svetlana Lazebnik
Understanding Scenes With Superpixels and Object Detectors
This talk will focus on some of my group's recent work on understanding scenes in terms of their constituent objects and global category-level properties.
First, I will discuss our approach to the problem of image parsing, i.e., segmenting images and labeling their regions with object classes. We have developed a nonparametric method that can accommodate "open universe" datasets where the training data and classes can evolve as users add images and annotations. Our method works by scene-level matching with global image descriptors, followed by superpixel-level matching with local features and MRF optimization for incorporating
neighborhood context. We have applied this system to a dataset of over 45K images from LabelMe with over 230 labels, and have also extended it to work on video. One of its biggest current shortcomings is its poor performance on rare categories and "things" (as opposed to "stuff"). I will discuss recent work on using per-exemplar detectors to overcome this limitation.
Finally, I will talk about another project on scene recognition using deformable part-based models. Even though DPM's have been introduced for fully supervised training of object detectors, we have found that they can also be effective for weakly supervised learning of visually consistent scene elements. By combining DPM's with standard global image features, we can obtain state-of-the-art results on the MIT 67-category indoor scene dataset.
Joint work with Joe Tighe and Megha Pandey.
Svetlana Lazebnik received her Ph.D. in 2006 at the University of Illinois at Urbana-Champaign. From 2007 to 2012, she was an assistant professor of computer science at the University of North Carolina at Chapel Hill. As of January 2012, she has moved back to UIUC as an assistant professor. She is the recipient of an NSF CAREER award and a Microsoft Research Faculty Fellowship, and is a member of the DARPA Computer Science Study Group. Her research interests include computer vision, image understanding, and machine learning techniques for visual data.