Abstract
One of the challenges in training graph neural networks (GNNs) applied to human-object interaction (HOI) is the computational complexity associated with updating and aggregating the information of all connected nodes in dense graph data, which results in a long training time and poor convergence efficiency. In particular, the parallel multi-head graph attention network (PMGAT), a graph neural network model, has achieved promising results in HOI detection by capturing the interactive associations between keypoints through local feature modules and multi-head graph attention mechanisms. However, to address the challenge of computational complexity, this study proposes a graph sampling-based dynamic edge strategy called GraphSADES to improve the PMGAT. GraphSADES reduces computational complexity by dynamically sampling a subset of edges during the training process while maintaining the precision of the original model. Initially, an object-centered complete graph is constructed, node updates are performed to obtain the initial attention coefficients, and importance coefficients are computed. Subsequently, a dynamic edge sampling strategy is adopted to reduce the computational complexity by randomly selecting a subset of edges for updating and aggregating the information in each training step. Through experimental comparative analysis, GraphSADES-PMGAT maintains the precision of the PMGAT model, and the models are trained using ResNet-50 and ViT-B/16 as backbone networks. On the dataset, HICO-DET, Floating Point Operations (FLOPs) for computational complexity are decreased by 40.12% and 39.89%, and the training time is decreased by 14.20% and 12.02%, respectively, and the convergence efficiency is the earliest to converge after 180 epochs. On the V-COCO dataset, under the same backbone network condition as HICO-DET, FLOPs decreased by 39.81% and 39.56%, training time decreased by 10.26% and 16.91%, respectively, and the convergence efficiency was the earliest to converge after 165 epochs. Specifically, GraphSADES-PMGAT maintains comparable precision while reducing FLOPs, resulting in a shorter training time and improved convergence efficiency compared to the PMGAT model. This work opens up new possibilities for achieving efficient human-object interaction detection.