Estimating a 2D pose from a tiny person image with super-resolution reconstruction. (July 2021)

• Record Type: Journal Article
• Authors: Zhang, Zhizhuo; Wan, Lili; Xu, Wanru; Wang, Shenghui
• Is Part Of: Computers & electrical engineering. Volume 93(2021)
• Publication Date: 2021-07
• Subjects: Human pose estimation -- Low-resolution problem -- Super-resolution reconstruction -- Pose-driven; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2021.107192
• Languages: English
• ISSNs: 0045-7906

Abstract

Abstract: Human pose estimation has achieved tremendous advances in accuracy with the emergence of various deep neural network architectures. However, for low-resolution (LR) images, we are far from achieving an acceptable accuracy. Deep learning based super-resolution (SR) has been proven to be helpful for addressing the challenges of face recognition and object detection in LR images. Following this idea, we integrate SR into existing human pose estimation networks to increase accuracy for LR images. In this work, we propose a novel end-to-end network architecture for the effective combination of SR and human pose estimation. Moreover, an approach is presented to guide the SR network to generate intermediate high-resolution (HR) images that contribute to pose estimation, rather than simply taking SR as an upstream task. The experimental results show that the accuracy of our approach has over 20% improved to that of the interpolation-assisted pose estimation network on the downsampled MPII dataset. Graphical abstract: Highlights: The features extracted from low-resolution images are very limited. Propose an architecture to recognize human poses in low-resolution images. Present a pose-driven method to recover images beneficial to pose estimation. Verify the effectiveness on downsampled MPII and Leeds Sports Pose datasets.