Hybrid Tow-Stream Information for Skeleton-Based Action Recognition Using Ensemble Learning

Skeleton data, a topological representation of the human body with joints and bones, are lightweight and compact, and thus are ideal candidates for action recognition. Ensemble learning has been shown to improve predictive accuracy or decompose a complex, difficult learning problem into more manageable sub-problems. However, studies have yet to attempt to utilize ensemble learning in GCN-based action recognition. This paper investigates the feasibility of ensemble learning on the GCN-based human action recognition task. Moreover, we start with data diversity aspects, i.e., first-order and second-order information, and utilize ensemble learning to improve the GCN model's recognition performance and generalization ability. First, we treat different GCN models as base models, such as ST-GCN, Js-AGCN, and Bs-AGCN, and extract the first-order and second-order information of the skeleton data using them. Then, two ensemble learning algorithms, i.e., Bagging and Blending algorithms, are utilized to ensemble the two-stream information. Experiments on the partial Kinetics-Skeleton dataset demonstrate that the performance of the ensemble model exceeds substantial improvements over the base models.