Self-calibrated stacked hourglass for human pose estimation

Abstract

In recent years, with the rapid development of deep learning algorithms such as CNN and RNN, human key point detection has been paid more and more attention. Human key point detection is always a research difficulty in computer vision because it is easily affected by complex joint occlusion, body deformation or lighting environment changes. With the proposed stacked hourglass network, it has become the mainstream method in the field of key point detection in recent years because of its repeatable upsampling and downsampling, and the addition of intermediate supervision model. Based on the stacked hourglass network, this paper introduces the self-calibrated convolutional network in SCNet network and optimates the structure of the original stacked hourglass network, and proposes a new human key point detection method named Self-calibrated stacked hourglass (SCHNs). This method effectively reduces the problem of inaccurate local feature description in the existing stacked hourglass network, and weakens the interference of useless signals in the global information to local information. Our method is trained and tested on the MPII human pose estimation dataset, and the case studys on the MSCOCO dataset. its detection results can effectively solve the problems existing in the existing stacked hourglass network, and it is significantly improved compared with other network models.