Can Super Resolution Improve Human Pose Estimation in Low Resolution Scenarios?

• DOI: 10.5220/0010863700003124
• Publication Date: 2022-02-14T15:57:35Z
• Authors: Peter Hardy; Srinandan Dasmaha...; Hansung Kim
• Keywords: Low resolution

The results obtained from state of the art human pose estimation (HPE) models degrade rapidly when evaluating people a low resolution, but can super resolution (SR) be used to help mitigate this effect? By using various SR approaches we enhanced two datasets and evaluated change in performance both an object keypoint detector as well end-to-end HPE results. We remark following observations. First find that for who were originally depicted at (segmentation area pixels), their detection would improve once was applied. Second, gained is dependent on persons pixel count original image prior any application SR; improved applied with small initial segmentation area, degrades becomes larger. To address introduced novel Mask-RCNN approach, utilising threshold decide use during step. This approach achieved best our each metrics.