Estimating a 2D pose from a tiny person image with super-resolution reconstruction

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.