MLP-Pose: Human Pose Estimation by MLP-Mixer

Current human pose estimation methods mainly use multi-scale fusion fully convolutional networks to achieve impressive results. However, this fully convolutional network lacks the ability to capture the relationship between features. In this paper, we propose a human pose estimation method based on MLP-Mixer. In detail, using 1D heatmaps as the ground truth, the human pose estimation is transformed into a sequence prediction problem on the horizontal axis and the vertical axis, so that the MLP-Mixer can be directly used to capture the relationship between the features. In addition, the existing backbone lacks intra-layer fusing. In order to solve this problem, we propose an efficient intra-layer fusion module. Specifically, our proposed MLP-Pose can achieve 77. 0AP and 76. 2AP on the COCO validation and test-dev dataset respectively.