The recent explosion in multimedia usage has resulted in a high demand for algorithms and applications to search large collections of images and videos. This technology describes the Columbia Transductive Annotation by Graph (TAG) system. TAG is designed to facilitate rapid retrieval and exploration of large image and video databases. It incorporates label propagation techniques and intuitive graphic user interfaces that allow users to quickly browse and annotate a small number of images and videos, then generate refined labels for all remaining data in the collection. This technology can be applied to domains such as intelligence, surveillance, consumer, biomedical, and web, providing a powerful tool to save search time and increase working efficacy.
While traditional approaches based on automatic image classification usually require a sufficiently large number of labeled samples to train classifiers (i.e. supervised learning), the TAG system minimizes the burden of manual labeling on users. The method is able to start with a small number of user-provided labels and propagate them to the remaining data to generate optimal ways of ranking, classifying, and presenting the multimedia data sets. The objective is to optimally leverage the available user input (as few as one or two samples per class) and propagate such information in the most effective way to all the remaining data in the collection.
Patent Pending (WO/2010/075408) Patent Pending (WO/2009/155281)
Tech Ventures Reference: IR M10-016