Pose estimation from Low-Resolution Images

Human pose estimation deals with inferring the spatial configuration of human bodies from images or videos. It is a fundamental task in computer vision engaged in the analysis of form and measurement of range of motion for the purpose of improving performance in sports, postural analysis for healthcare such as rehabilitation and improvement of postural distortion, detection of characteristic movements from images from surveillance cameras, and detection of trespassing. Most studies involving deep learning for pose estimation assume the input images are clear and of a high quality. However, we often come across old and blurred images from which application programs need to extract and estimate the pose of individuals in real life systems. This study aims to observe the effects of low image quality on pose estimation, and to explore methods that can demonstrate high accuracy even in low-quality images. A modified and fine-tuned ResNet 50 model is trained on the MPII Human Pose Dataset containing 22,246 images after lowering their resolution. Training consists providing an input image of a person and estimating the coordinates of the joint position with as little error as possible in the test. The PCKh@0.5 averages show a difference in accuracy with decreasing image quality as expected. However, the overall results demonstrate that pose estimation can be performed while maintaining a certain degree of accuracy even under low image quality.