L-HRNet: A Lightweight High-Resolution Network for Human Pose Estimation

Aiming at the problem that the human pose estimation network model is trying to improve the accuracy and thus leads to a large number of parameters and computations, this paper proposes a lightweight human pose estimation network model. We present a solution called L-HRNet that incorporates Depthwise Separable Convolution, Sandglass module, and Attention Mechanism to reduce the network's parameters and computational complexity while maintaining a high level of accuracy. Compared to the high-resolution network (HRNet [10]), L-HRNet's model size (#Params) is only 5.6%, and computational complexity (FLOPs) is only 11.9%. Our L-HRNet demonstrates both effectiveness and efficiency on a benchmark dataset: COCO keypoint detection dataset, achieving 65.2 AP on the COCO test-dev set with only 1.59 M parameters and 0.89 GFLOPs. The code and models are publicly available at https://github.com/ApingJJ/L-HRNet.git.