A large convolution model with large receptive field for human pose estimation

Inhalt:
We propose a large convolution model to complete the human pose estimation task, which has a large effective receptive field and reduces the destruction of channel information by the ReLU function. We adjusted the model of lightweight Openpose and used the structure of mobilenetv3 as the backbone of lightweight Openpose, hoping to use the InvertedResidual structure to reduces the destruction of channel information by the ReLU function. Insert RepLKBlock into the backbone to improve the effective receptive field of the model. Our model improves the following problems of the lightweight Openpose model. Lightweight Openpose has a small effective receptive field. The destruction of information by the ReLU function of lightweight Openpose. Large convolution model achieves Average Precision (AP) of 0.47 on the COCO2017 validation set.