Semi-supervised Learning Combining 2DCNNs and Video Compression for Action Recognition
Hayato Terao, Wataru Noguchi, Hiroyuki Iizuka, Masahito Yamamoto · 2020
Action recognition is a classification task of actions in videos. Recently, 3DCNNs are widely used for action recognition. However, a large annotated dataset is necessary for achieving high performances with 3DCNNs, and the performance is often limited when it is hard to prepare a sufficient amount of annotated data. Semi-supervised learning is an approach to overcome this problem, but a few works pay attention to semi-supervised action recognition. Semi-supervised learning is actively studied in image classification. We can apply these semi-supervised learning methods directly to train 3DCNNs for action recognition. However, Jing et al. show that this approach does not significantly improve performance.