Leveraging Pre-trained CNN Models for Skeleton-Based Action Recognition

Skeleton-based human action recognition has recently drawn increasing attention thanks to the availability of low-cost motion capture devices, and accessibility of large-scale 3D skeleton datasets. One of the key challenges in action recognition lies in the high dimensionality of the captured data. In recent works, researchers draw inspiration from the success of deep learning in computer vision in order to improve the performances of action recognition systems. Unfortunately, most of these studies do not leverage different available deep architectures but develop new architectures. Most of the available architecture achieve very high accuracy in different image classification problems. In this paper, we use these architectures that are already pre-trained on other image classification tasks. Skeleton sequences are first transformed into image-like data representation. The resulting images are used to train different state-of-the-art CNN architectures following different training procedures. The experimental results obtained on the popular NTU RGB+D dataset, are very promising and outperform most of the state-of-the-art results.