Recognition of rare low-moral actions using depth data

Abstract

Detecting and recognizing low-moral actions in public spaces is important. But low-moral actions are rare, so in order to learn to recognize a new low-moral action in general we need to rely on a limited number of samples. In order to study the recognition of actions from a comparatively small dataset, in this work we introduced a new dataset of human actions consisting in large part of low-moral behaviors. In addition, we used this dataset to test the performance of a number of classifiers, which used either depth data or extracted skeletons. The results show that both depth data and skeleton based classifiers were able to achieve similar classification accuracy on this dataset (Top-1: around 55%, Top-5: around 90%). Also, using transfer learning in both cases improved the performance.