LW-FastPose: A Lightweight Network for Human Pose Estimation Based on Improvements to FastPose
Mingxing Cai, Wang‐Su Jeon, Sang‐Yong Rhee · 2025
As human pose estimation network models continue to improve in performance, deeper network structures often come with a significant increase in parameters and computational complexity. To address this issue, this paper proposes a lightweight human pose estimation network, LW-FastPose. The network is composed of an EL-ResNet feature extraction network and an upsampling DUC module. The foundation of the EL-ResNet feature extraction network, the DWNK base module, is a depthwise separable convolution module that leverages residual connections and depthwise separable convolution techniques. This design significantly reduces computational complexity while maintaining model performance. Experimental results on the COCO dataset show that compared to the FastPose network, LW-FastPose reduces the number of parameters and computational complexity by 34% and 27%, respectively. Compared to other mainstream models, LW-FastPose not only significantly reduces parameter count and computational load but also achieves competitive prediction accuracy, without compromising performance.