A High-Resolution Human Pose Estimation Method with Coordinate Attention
Zhengjie Deng, Yuanzhong Li, Shuqian He, Yizhen Wang, Xun Wang · 2022
Human pose estimation is an important element in video and image analysis. However, existing methods have some problems such as less accurate predicted position and over-lapping key points. To address these problems, we construct a high-resolution human pose estimation method based on Coordinate Attention (CA) mechanism that combines coordinate perception. The method constructs a coordinate attention bottleneck module (CABottleneck) and a coordinate attention module (CABlock) based on the attention mechanism, which allow the model to obtain richer coordinate and orientation information in feature extraction, and then combine the downsampling operation to convert the original image into four different sizes of feature maps, and then fuse these feature maps in multivariate resolution. Then, we use the deconvolution module to obtain the high-resolution feature maps, and finally derive the positions of each key point to achieve the estimation of human pose. The experimental results show that the accuracy of the method in this paper is better than that of various methods such as HigherHRNet when using similar number of parameters. In the CrowdPose test dataset, the accuracy of both single-scale and multi-scale exceeds 72%, which is 6.6% better than HigherHRNet; 5% better than DEKR. In the COCO2017 val dataset, our method maintains a high performance and reduces the repetition rate of human key points.