A high precision resolution network MGS-HigherHRnet for human pose estimation

Bottom-up human pose estimation models perform well in complex scenarios, but such models are complex and computationally expensive. To address the above problems, this paper proposes a bottom-up human pose estimation network model MGS-HigherHRnet based on HigherHRnet. The MGS-HigherHRnet was transformed from the HigherHRnet. The standard convolution unit in Bottleneck module and Basicblock module of the HighrtHRnet model was replaced by MBConv unit and GhostConv unit, respectively. After the replacement of Bottleneck module and Basicblock module, they were called MGSneck moudule and MGSblock module respectively in this paper. Furthermore, the SE attention mechanism unit has been added to both MGSneck module and MGSblock module. This replacement utilized of the high efficiency of the Ghost convolution unit in generating redundant features. It also leveraged the low computational effort of the MBConv depth-separable convolutions. Additionally, feature extraction was enhanced by the SE attention mechanism. As a result, it effectively reduced the computational complexity while maintaining the overall prediction performance of the model. The model in this paper was validated on the COCO dataset. The experimental results showed that, compared to HigherHRNet, the MGS HigherHRnet network model not only reduced computational complexity slightly, but also effectively improves prediction accuracy. The parameter count of MGS-HigherHRnet is basically the same as that of HigherHRnet. The computational cost of MGS-HigherHRnet is reduced by 14.8%. The prediction accuracy of key points has improved by 1.9 pp. Especially, the prediction accuracy of key points in large scale persons improved by 3.9 pp.