Feature Learning Capacity Assessment of Deep Convolutional Generative Adversarial Network for Action Recognition in a Self-Supervised Framework

Feature learning has always been a critical and most important problem in the field of computer vision. Most of the research community is addressing the problem of feature learning using supervised learning which requires a lot of manually annotated data. In this paper, a self-supervised framework is proposed to evaluate the feature learning capability of the discriminator of a deep convolutional generative adversarial network (DCGAN) via action classification. The DCGAN is trained on action videos of the UCF101 dataset without using any label information and then the trained discriminator is extracted from the DCGAN network. The trained discriminator is used to generate feature vectors. The action classification is performed by finding the similarity between these feature vectors using multiple similarity measures. The experimental results prove that discriminator is a good feature vector generator as the maximum number of action classes are classified correctly without using any annotated data.