Hardmining Training Via Self-Adversarial Network for Human Pose Estimation

Abstract

Human pose estimation, as a branch of machine vision, has broad application prospects in the fields of behavior recognition, human-computer interaction and so on. Although the current human pose estimation method has made good progress, there is still room for improvement in the prediction of difficult joint points. In this paper, the hardmining training technique is proposed to improve this problem. We use self-adversarial network as our training model, which consists of two stacked hourglasses with the same architecture, one as the generator and the other as the discriminator. During the training period, the discriminator distinguishes the generated heatmaps from the ground-truth heatmaps, and introduces the adversarial loss to the generator through back-propagation to induced generator generates more reasonable prediction, on this basis, we introduce a method called hardmining to focus the training attention on the difficult joint points, thus improving the prediction accuracy on difficult joint points. After the training is done, the generator is used as a human pose estimator.