Action Recognition with Joints-Pooled 3D Deep Convolutional Descriptors

Torso joints can be considered as the landmarks of human body. An action consists of a series of body poses which are determined by the positions of the joints. With the rapid development of RGB-D camera technique and pose estimation research, the acquisition of the body joints has become much easier than before. Thus, we propose to incorporate joint positions with currently popular deep-learned features for action recognition. In this paper, we present a simple, yet effective method to aggregate convolutional activations of a 3D deep convolutional neural network (3D CNN) into discriminative descriptors based on joint positions. Two pooling schemes for mapping body joints into convolutional feature maps are discussed. The joints-pooled 3D deep convolutional descriptors (JDDs) are more effective and robust than the original 3D CNN features and other competing features. We evaluate the proposed descriptors on recognizing both short actions and complex activities. Experimental results on real-world datasets show that our method generates promising results, outperforming state-of-the-art results significantly.