Understanding actions in videos remains a significant challenge in computer vision, which has been the subject of several pieces of research in the last decades. Convolutional neural networks (CNN) are a significant component of this topic and play a crucial role in the renown of Deep Learning. Inspired by the human vision system, CNN has been applied to visual data exploitation and has solved various challenges in various computer vision tasks and video/image analysis, including action recognition (AR). However, not long ago, along with the achievement of the transformer in natural language processing (NLP), it began to set new trends in vision tasks, which has created a discussion around whether the Vision Transformer models (ViT) will replace CNN in action recognition in video clips. This paper conducts this trending topic in detail, the study of CNN and Transformer for Action Recognition separately and a comparative study of the accuracy-complexity trade-off. Finally, based on the performance analysis’s outcome, the question of whether CNN or Vision Transformers will win the race will be discussed.

3D-Skeleton-based action recognition has been widely adopted due to its efficiency and robustness to complex backgrounds. While it is capable of conveying a significant amount of information regarding the dynamics of human poses, we argue that its performance is curtailed when confronted with actions involving interactions between humans and objects due to the absence of the study of the surrounding objects. It is of great importance to delve deeper into the study of human-object interactions for skeleton-based action recognition. This paper proposes a novel approach to represent the spatial-temporal skeleton features, along with the present nearby objects and their dynamics. To accomplish this, a new formulation named object knowledge is presented, which entails the categorization of object characteristics, based on whether or not the object necessitates a motion analysis. With a piece of prior knowledge, in cases where it is required, the motion is calculated, while for those where it is not necessary, only the category of object is considered. This object knowledge is then early-fusion along with the skeleton representation, in such a way that it fits into the self-attention model. The experimental results on different popular action recognition datasets (NTU RGB+D 60, NTU RGB+ D 120 illustrate that the proposed approach outperforms the current state-of-the-art methods.