Home | Search | Help  
Home Page Università di Genova

Seminar Details


Date 18-3-2008
Time 14:30
Room/Location Sala Conferenze- DISI - 3 piano
Title Large scale visual concept learning
Speaker Jan-Mark Geusebroek
Affiliation Intelligent Systems Lab Amsterdam, University of Amsterdam, The Netherlands
Link http://slipguru.disi.unige.it/Events/seminars.php
Abstract Visual categorization is a basic means to process and organize vast amounts of image data. For humans, it provides cognitive control over the environment by memorizing, reasoning, and learning to learn. However, the automatic learning of many concepts from category-labeled images is non trivial. Initial approaches to scene and object categorization for a small number of categories are promising, but how can we scale this up to thousands of categories? In this talk, I will present some of our recent results on large scale concept learning in image and video collections. Scaling to many categories can only be achieved with compact yet expressive visual models. I will first briefly describe how to construct compact color image descriptors using natural image statistics. I will show how this in turn can be used in visual word based scene categorization. Next, I will describe how to construct visual models to facilitate large scale video indexing. I will show how the inclusion of ambiguity in the codebook model further improves expressiveness. Furthermore, I show three methods that allow compact, yet discriminative vocabularies. I will demonstrate the effectiveness of the methods by results in video retrieval for TRECVID, for scene categorization on standard datasets, and for rough scene geometry estimation from single images.
Back to Seminars