A simple baseline for domain generalization of action recognition and a realistic out-of-domain scenario

In this study, we tackle the Domain Generalization (DG) challenge in the field of video action recognition. Previous works on DG for action recognition mostly focus on the Single Domain Generalization (SDG) problem, which is difficult to achieve due to domain biases present in datasets. Thus, we aim to address this challenge by re-defining DG for video action recognition. Another limitation of existing research is the lack of realistic out-of-domain scenarios. To overcome this, we introduce a new benchmark for DG in video action recognition named ENT dataset that contains realistic domain shifts. Additionally, we propose a straightforward baseline for DG in the video domain that helps improve the performance of action recognition models in unseen domains. Our proposed approach adapts the image-based methods to the video domain, leading to improved DG performance.