Lightweight human pose estimation based on dynamic ghost
To address the issue of the inadequate accuracy of current lightweight human pose estimation network when detecting under reduced parameter count and computational complexity,a lightweight human pose estimation network based on dynamic ghost(dynamic ghost network,DGNet)was proposed.DGNet employs an innovative approach to succinctly and effectively extract con-textual information,enhancing the model's representation capability and consequently improving performance without increasing parameter count and computational complexity.Specifically,the model utilizes dynamic shuffling and ghost operations to construct two novel lightweight modules:the dynamic ghost neck module(DGNeck)and the dynamic ghost basicblock module(DGBlock).DGNeck replaces convolution operations with less costly linear operations to reduce network parameters and computational complex-ity.Simultaneously,DGBlock dynamically aggregates multiple channels and shuffles them to obtain accurate positional information in the feature map,thus improving detection accuracy.Experimental results under comparable conditions show that,compared to existing Lite-HRNet models,DGNet model achieves a 4.8%reduction in computational complexity and a 2.3%elevation in accu-racy on the COCO validation set,while on the MPII validation set,it achieves a 3.7%reduction in computational complexity and a 0.7%increase in accuracy.
human pose estimationdeep neural networkhigh-resolution networklightweight
NETL
NSTL
万方数据
