Image-based action recognition using hint-enhanced deep neural networks

While human action recognition from still images finds wide applications in computer vision, it remains a very challenging problem. Compared with video-based ones, image-based action representation and recognition are impossible to access the motion cues of action, which largely increases the difficulties in dealing with pose variances and cluttered backgrounds. Motivated by the recent success of convolutional neural networks (CNN) in learning discriminative features from objects in the presence of variations and backgrounds, in this paper, we investigate the potentials of CNN in image-based action recognition. A new action recognition method is proposed by implicitly integrating pose hints into the CNN framework, i.e., we use a CNN originally learned for object recognition as a base network and then transfer it to action recognition by training the base network jointly with inference of poses. Such a joint training scheme can guide the network towards pose inference and meanwhile prevent the unrelated knowledge inherited from the base network. For further performance improvement, the training data is augmented by enriching the pose-related samples. The experimental results on three benchmark datasets have demonstrated the effectiveness of our method.