Video retrieval based on object discovery

State-of-the-art video retrieval methods use global image statistics to provide low level descriptors or use object recognizers to provide high level features. Using global image statistics can be hindered by lack of explicitly characterizing the object of interest hence prone to retrieving irrelevant results, while using object recognizers can suffer from having to train a large number of object recognizers for different types of objects. We present a novel framework for content-based video retrieval. We use an Unsupervised learning method to automatically discover and locate the object of interest in a video clip. This Unsupervised learning algorithm alleviates the need for training a large number of object recognizers. Regional image characteristics are extracted from the object of interest to form a set of descriptors for each video. A novel ensemble-based matching algorithm compares the similarity between two videos based on the set of descriptors each video contains. Videos containing large pose, size, and lighting variations are used to validate our approach. (C) 2008 Elsevier Inc. All rights reserved.