Internal Transfer Learning for Improving Performance in Human Action Recognition for Small Datasets

Human action recognition nowadays plays a key role in varieties of computer vision applications. Many computer vision methods focus on algorithms designing classifiers with handcrafted features which are complex and inflexible. In this paper, we focus on the human action recognition problem and utilize 3D convolutional neural networks to automatically extract both spatial and temporal features for classification. Specifically, in order to address the training problems with small data sets, we propose an internal transfer learning strategy adapted to this framework, by incorporating the sub-data classification method into transfer learning. We evaluate our method on several data sets and obtain promising results. With the proposed strategy, the performance of human action recognition is improved obviously.