Learning representative temporal features for action recognition

In this paper, a novel video classification method is presented that aims to recognize different categories of third-person videos efficiently. Our motivation achieve light model could be trained with insufficient training data. With intuition, the processing 3-dimensional input broken 1D in temporal dimension on top 2D spatial. The processes related spatial frames are being done by utilizing pre-trained networks no phase. only step which involves classify time series resulted from description signals. As matter fact, optical flow images first calculated consecutive and described CNN networks. Their then reduced using PCA. By stacking vectors beside each other, multi-channel created for video. Each channel represents specific feature follows it over time. main focus proposed obtained effectively. Towards this, idea let machine learn features. This one dimensional Convolutional Neural Network (1D-CNN). 1D-CNN learns features along dimension. Hence, number parameters decreases significantly would result trainability even smaller datasets. It illustrated reach state-of-the-art results two public datasets UCF11, jHMDB competitive HMDB51.