A new architecture of neural network for fine-grained video analysis based on visual attention
Lin Li, SUN Kangbo, ZHU Jie · DOAJ (DOAJ: Directory of Open Access Journals) · 2019
Based on a subtle gesture video dataset,a deep network was proposed to improve the performance on such fine-grained actions in this paper.The architecture consisted of a smaller variation of convolutional 3-dimensional (C3D) network and a long short-term memory (LSTM) with a soft attention mechanism.The depth of C3D network and the weight penalty of the attention mechanism were optimized.Experimental results showed that the fine-grained action recognition network could effectively focus on the certain part with important information,and performed better on both average accuracy and detection accuracy for the dataset.