Video multitask transformer network

In this paper, we propose the Multitask Transformer Network for multitasking on untrimmed video. To analyze the untrimmed video, it needs to capture important frame and region in the spatio-temporal domain. Therefore, we utilize the Transformer Network, which can capture the useful features from CNN representations through an attention mechanism. Motivated by the Action Transformer Network, which is a repurposed model of the Transformer for video, we modified the concept of query which was specialized only for action recognition on the trimmed video to fit the untrimmed video. In addition, we modified the structure of the Transformer unit to the pre-activation structure for identity mapping on residual connections. We also utilize the class conversion matrix (CCM), one of the feature fusion methods, to share the information of different tasks. Combining our Transformer structure and CCM, the Multitask Transformer Network is proposed for multitasking on untrimmed video. Eventually, our model evaluated on CoVieW 2019, and we enhanced the performance through post-processing based on prediction results that suitable to the CoVieW 2019 evaluation metric. In CoVieW 2019 challenge, we placed fourth on final rank while first on scene and action score.