Understanding the limits of 2D skeletons for action recognition

With the development of motion capture technologies, 3D action recognition has become a popular task that finds great applicability in many areas, such as augmented reality, human–computer interaction, sports, or healthcare. On the other hand, the acquisition of 3D human skeleton data is an expensive and time-consuming process, mainly due to the high costs of capturing technologies and the absence of suitable actors. We overcome these issues by focusing on the 2D skeleton modality that can be easily extracted from ordinary videos. The objective of this work is to demonstrate a high descriptive power of such a 2D skeleton modality by achieving accuracy on the task of daily action recognition competitive to 3D skeleton data. More importantly, we thoroughly analyze the factors that significantly influence the 2D recognition accuracy, such as the sensitivity towards data normalization, scaling, quantization, and 3D-to-2D distortions in skeleton orientations and sizes, which are caused by the loss of depth dimension and fixed-angle camera view. We also provide valuable insights on how to mitigate these problems to increase recognition accuracy significantly. The experimental evaluation is conducted on three datasets different in nature. The ability to learn different types of actions better using either 2D or 3D skeletons is also reported. Throughout experiments, a generic light-weight LSTM network is used, whose architecture can be easily tuned to achieve the desired trade-off between its accuracy and efficiency. We show that the proposed approach achieves not only the state-of-the-art results in 2D skeleton action recognition but is also highly competitive to the best-performing methods classifying 3D skeleton sequences or the visual content extracted from ordinary videos.