Toward Robust Human Pose Estimation under Real-World Image Degradations and Restoration Scenarios

Human Pose Estimation (HPE) models have varied applications and represent a cutting-edge branch of study, whose systems such as MediaPipe (MP) and OpenPose (OP) showing marked success. One of these areas, however, that is inadequately researched is the impact of image degradation on the accuracy of HPE models. Image degradation refers to images whose visual quality has been purposefully diminished by means of techniques like brightness adjustments- which can lead to an increase or a decrease in the intensity levels-, geometric rotations, or resolution downscaling. The study of how these types of degradation impact the functionality of HPE models is a virtually unexplored area. In addition, current methods of improving degraded images to a high quality have not been well examined in relation to improving HPE models. In this study, we clearly demonstrate a drop in the precision of the HPE model when image quality is degraded. Our qualitative and quantitative measurements identify a wide difference in performance in identifying landmarks as images undergo changes in brightness, rotations, or reductions in resolution. Additionally, we have tested a variety of existing image enhancement methods in an attempt to enhance their capability in restoring low quality images, hence supporting improved functionality of HPE. Interestingly, for rotated images, using Pillow of OpenCV improves landmark recognition precision drastically, nearly restoring it to levels we see in high-quality images. In instances of brightness variation and in low-quality images, however, existing methods of enhancement fail to yield the improvements anticipated, highlighting a large direction of study that calls for additional research. In this regard, we proposed a wide-ranging system for classifying different types of image degradation systematically and for selecting appropriate algorithms for image restoration, in an effort to restore image quality. In a relative study of current methods, the Tuned RotNet model surpass official RotNet model in predicting rotation degree of images, where the accuracy of official RotNet and Tuned RotNet classifier were of 61.59\% and 92.04\%, respectively. Besides, in an effort to make it easier for other studies, we provide a new dataset of reference images and corresponding degenerated images, since currently there is a lack of controlled comparatives.