On the Importance of Temporal Features in Domain Adaptation Methods for Action Recognition

One of the most common vision problems is Video based Action Recognition. Many public datasets, public contests, and so on, boosted the development of new methods to face the challenges posed by this problem. Deep Learning is by far the most used technique to address Video-based Action Recognition problem. The common issue for these methods is the well-known dependency from training data. Methods are effective when training and test data are extracted from the same distribution. However, in real situations, this is not always the case. When test data has a different distribution than training one, methods result in considerable drop in performances. A solution to this issue is the so-called Domain Adaptation technique, whose goal is to construct methods that adapt test data to the original distribution used in training phase in order to perform well on a different but related target domain. Inspired by some existing approaches in the scientific literature, we proposed a modification of a Domain Adaptation architecture, that is more efficient than existing ones, because it improves the temporal dynamics alignment between source and target data. Experiments show this performance improvement on public standard benchmarks for Action Recognition