Low-Resolution Human Pose Estimation and Action Recognition via Pose-Driven Super-Resolution Reconstruction

Abstract

In recent years, human pose estimation quality has been greatly improved by deep learning. However, for a tiny human image, the limited information carried by the low-resolution image brings robustness issues to human pose estimation. To increase the amount of information, we introduce super-resolution (SR) reconstruction into human pose estimation. We propose a novel low-resolution human pose estimation method, which effectively combines SR reconstruction and human pose estimation. Different from other SR reconstruction algorithms, our pose-driven SR reconstruction is guided to generate intermediate results conducive to human pose estimation. Moreover, considering that good pose estimation results are crucial to pose-related action recognition, we present a low-resolution human action recognition solution that applies our pose estimation method to pose-related action recognition. Experimental results show that our method can significantly improve the performance of the existing pose estimation and action recognition networks when processing low-resolution images.