| Summary: | This work develops Deep Neural Networks (DNNs) by adopting Capsule Networks (CapsNets) and spatiotemporal skeleton-based attention to effectively recognize subject actions from abundant spatial and temporal contexts of videos. The proposed generic DNN includes four 3D Convolutional Neural Networks (3D_CNNs), Attention-Jointed Appearance (AJA) and Attention-Jointed Motion (AJM) generation layers, two Reduction Layers (RLs), two Attention-based Recurrent Neural Networks (A_RNNs), and an inference classifier, where RGB, transformed skeleton, and optical-flow channel streams are inputs. The AJA and AJM generation layers emphasize skeletons to the appearances and motions of a subject, respectively. A_RNNs generate attention weights over time steps to highlight rich temporal contexts. To integrate CapsNets in this generic DNN, three types of CapsNet-based DNNs are devised, where the CapsNets take over a classifier, A_RNN+classifier, and RL+A_RNN+classifier. The experimental results reveal that the proposed DNN using CapsNet as an inference classifier outperforms the other two CapsNet-based DNNs and the generic DNN adopting the feedforward neural network as an inference classifier. Additionally, our best CapsNet-based DNN achieves average accuracies of 98.5% for the state-of-the-art performance in UCF101, 82.1% for near-state-of-the-art performance in HMDB51, and 95.3% for panoramic videos, to the best of our knowledge. Particularly, it is determined that the generic CapsNet behaves as an outstanding inference classifier but is slightly worse than the A_RNN in interpreting temporal evidence for recognition. Therefore, the proposed DNN, which employs CapsNet to fulfill an inference classifier, can be superiorly applied to various context-aware visual applications.
|
|---|