PyCaNet: Pose Estimation of Primary and Middle School Students Using Pyramid Convolutions and Coordinate Attention

Abstract 2D pose estimation task is the basics of action recognition, human-computer interaction and other applications, but there is few research on pose estimation of primary and middle school students. In this paper, we present APMSS: a new dataset based on primary and middle school scenarios. The dataset fills the gap of the current human pose estimation dataset with few campus scenes, and the human keypoints and object bounding boxes in the dataset are manually annotated. In addition, we propose PyCaNet: a new top-down human pose estimation neural network model that uses pyramid convolutions to give the network a larger receptive field for better results. Coordinate attention is also added to the network, which is a kind of channel attention and can significantly improve the network performance. Extensive experiments shows that PyCaNet outperforms the previous methods by at least 1.0 AP on APMSS val and at least 0.7 PCKh on MPII test.