LCMV: Lightweight Classification Module for Video Domain Adaptation

Video action recognition models exhibit high performance on in-distribution data but struggle with distribution shifts in test data. To mitigate this issue, Unsupervised Domain Adaptation (UDA) methods have been proposed, consisting in training on labeled data from a source domain and incorporating unlabeled test data from a target domain to reduce the domain gap. This requires simultaneous access to data from both domains, which may not be practical in real-world scenarios due to privacy issues. A more practical approach called Source-Free Domain Adaptation (SFDA) has been recently proposed, which consists in adapting a well-trained source model using only unlabeled target data. However, existing SFDA methods are computationally intensive and designed for specific architectures. In this paper, we propose an approach called Lightweight Classification Module for Video Domain Adaptation (LCMV). LCMV is based on a backpropagation-free prototypical algorithm, which efficiently adapts a source model using unlabeled target data only. Results on two popular datasets, HMDB-UCFfull and EPIC-Kitchens-55, show significant improvements of LCMV compared to the previous state-of-the-art SFDA methods, and competitive results when compared to state-of-the-art UDA methods.