Probability Matrix SVM+ Learning for Complex Action Recognition

Abstract

Complex action recognition is a hot topic in computer vision. When training a robust model, a large amount of labeled data is required. However, labeling complex actions is often time-consuming and expensive. Considering that each complex action is composed of a sequence of simple actions, we propose a new perspective to provide more information during training in order to solve the problem of insufficient labeled data. The probability matrix is then designed by manual annotation, which encodes a probability distribution of simple actions in complex actions. So the probability matrix is only available during training but unavailable during testing. Finally, a probability matrix is regared as privileged information in a SVM+ framework, and we regard this setting as probability matrix SVM+(pmSVM+). To validate the proposed model, extensive experiments are carried out on complex action datasets. Experiment results show the effectiveness of pmSVM+ for complex action recognition.