Statistical Validation for Automatic Feature Selection Algorithms in Video Applications

摘要

High time complexity is a bottle-neck in video segmentation, classification and retrieval. Automatic feature selection algorithms can speedup video analysis for more efficient and effective video data management of varied applications. Heuristic approaches are necessary because of the difficulty of seeking optimal solution, especially for high dimensional feature space and complex hypothesis space. This paper addresses the absence of the theoretic support for these heuristic feature selection algorithms in video applications. We propose four artificial datasets to simulate the behaviors of automatic feature selection in video domain. Different feature selection algorithms have been compared for accuracy and efficiency under different distributions of relevant features, different sizes of original feature spaces, different sizes of target feature spaces, different methods for classification and induction. Robustness of feature selection algorithms with sparse and noisy training data has also been discussed.