An adversarial semi-supervised approach for action recognition from pose information

The collection of video data for action recognition is very susceptible to measurement bias; the equipment used, camera angle and environmental conditions are all factors that majorly affect the distribution of the collected dataset. Inevitably, training a classifier that can successfully generalize to new data becomes a very hard problem, since it is impossible to gather general enough training sets. Recent approaches in the literature attempt to solve this problem by augmenting a given training set, with synthetic data, so as to better represent the global distribution of the covariates. However, these approaches are limited because they essentially involve hand-crafted data synthesizers, which are typically hard to implement and problem specific. In this work, we propose a different approach to tackling the above issues, which relies on the combination of two techniques: pose extraction, and domain adaptation as a means to improve the generalization capabilities of classifiers. We show that adapted skeletal representations can be retrieved automatically in a semi-supervised setting and these help to generalize classifiers to new forms of measurement bias. We empirically validate our approach for generalizing across different camera angles.