Video frame categorization using sort-merge feature selection

Feature selection for video categorization is impractical with existing techniques. We present a novel algorithm to select a very small subset of image features. We reduce the cardinality of the input data by sorting the individual features by their effectiveness in categorization, and then merging pairwise these features into feature sets of cardinality two. Repeating this sort-merge process several times results in the learning of a small-cardinality, efficient, but highly accurate, feature set. The cost of this wrapper method for learning the feature set, approximately O(F logF) where F is the number of incoming features, is very reasonable, particularly when compared with the impracticality of applying the much higher cost current filter or wrapper learning models to the massive data of this domain. We provide empirical validation of this method, comparing it to both random and hand-selected feature sets of comparable small cardinality.