查看文章

Convolutional neural networks have achieved great success in many computer vision tasks. However, it is still challenging for action recognition in videos due to the intrinsically complicated space-time correlation and computational difficult of videos. Existing methods usually neglect the fusion of long term spatio-temporal information. In this paper, we propose a novel hybrid spatio-temporal convolutional network for action recognition. Specifically, we integrate three different type of streams into the network: (1) the image stream utilizes still images to learn the appearance information; (2) the optical stream captures the motion information from optical flow frames; (3) the dynamic image stream explores the appearance information and motion information simultaneously from generated dynamic images. Finally, a weighted fusion strategy at the softmax layer is utilized to make the class decision. With the help of …