Motion based foreground detection and poselet motion features for action recognition

For action recognition, the actor(s) and the tools they use as well as their motion are of central importance. In this paper, we propose separating foreground items of an action from the background on the basis of motion cues. As a consequence, separate descriptors can be defined for the foreground regions, while combined foreground-background descriptors still capture the context of an action. Also a low-dimensional global camera motion descriptor can be computed. Poselet activations in the foreground area indicate the actor and its pose. We propose tracking these poselets to obtain detailed motion features of the actor. Experiments on the Hollywood2 dataset show that foreground-background separation and the poselet motion features lead to consistently favorable results, both relative to the baseline and in comparison to the current state-of-the-art.