Lightweight Non-local High-Resolution Networks for Human Pose Estimation

Abstract

Human pose estimation is one of the fundamental tasks in computer vision, applied in areas such as motion recognition, games, and animation production. Most of the current deep network models entail deepening the number of network layers to obtain better performance. This requires computational resources that exceed the computational capacity of embedded and mobile devices, thereby limiting the practical application of these approaches. In this paper, we propose a lightweight network model that incorporates the idea of Ghost modules. We design Ghost modules to replace the base modules in the original high-resolution network, thus reducing the network model parameters. In addition, we design a non-local high-resolution network that is fused in the 1/32 resolution stage of the network. This enables the network to acquire global features, thus improving the accuracy of human pose estimation and reducing the network parameters while ensuring the accuracy of the model. We verify the algorithm on the MPII and COCO datasets and the proposed model achieves a 1.8% improvement in accuracy while using 40% fewer parameters compared with the conventional high-resolution network.