MSRDN: A Super-Resolution Network for Human Body

Recent advances in human parsing estimation implicitly assume the input image is of sufficiently high resolution (HR). However, person images are everywhere especially in the surveillance applications. This paper aims at super resolving a person image with different resolution. First of all, a multi-scale residual dense block (MSRDB) is proposed to completely gather sufficient features for human body image super resolution (SR) task. And then, a multi-scale residual dense network (MSRDN) is developed, which consists of a cascaded MSRDBs, to reconstruct HR outputs in a way that goes from coarse to fine. Experiments on human body benchmark datasets show that our method performs better than other SOTA methods in terms of PSNR and SSIM metrics. Further, body landmark estimation task is introduced as novel evaluation metrics for human body. Performance on the task further proves the effectiveness of our method.