Fast Video Retrieval under Sparse Training Data

Abstract

Feature selection for video retrieval applications is impractical with existing techniques, because of their high time complexity and their failure on the relatively sparse training data that is available given video data size. In this paper we present a novel heuristic method for selecting image features for video, called the Complement Sort-Merge Tree (CSMT). It combines the virtues of a wrapper model approach for better accuracy with those of a filter method approach for incrementally deriving the appropriate features quickly. A novel combination of Fastmap for dimensionality reduction and Mahalanobis distance for likelihood determination is used as the induction algorithm. The time cost of CSMT is linear in the number of features and in the size of the training set, which is very reasonable. We apply CSMT to the domain of fast video retrieval of extended (75 minutes) instructional videos, and demonstrate its high accuracy in classifying frames.