基于动态幻影的轻量级人体姿态估计

Lightweight human pose estimation based on dynamic ghost

针对当前轻量级人体姿态估计网络在减少参数量和计算复杂度时未能有效提高检测精度的问题,提出了基于动态幻影的轻量级人体姿态估计网络(dynamic ghost network,DGNet)。DGNet采用一种创新的方法,能够简洁有效地提取上下文信息,实现在不增加参数量和计算复杂度的情况下提高模型的表征能力进而提升性能。具体而言,模型使用动态混洗和幻影操作构建2个全新的轻量级模块——动态幻影瓶颈模块(dynamic ghost neck module,DGNeck)和动态幻影基础模块(dynamic ghost basicblock module,DGBlock)。DGNeck将卷积运算替换为代价较小的线性运算进而降低网络参数和计算复杂度,同时DGBlock动态聚合多个通道并混洗,获取特征图精确位置信息以提高检测精度。同等条件下的实验结果表明,与现有的Lite-HRNet模型相比,DGNet模型在COCO校验集上计算复杂度下降了4.8%,准确率提高了2.3%,而在MPII校验集上计算复杂度降低了3.7%,准确率提高了0.7%。

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.