DE-HRNet: Detail enhanced high-resolution network for human pose estimation

Scale variation is a challenge in human pose estimation. The scale variations of human body are related to the accuracy and robustness of posture estimation. For example, the prediction accuracy of smaller joints (such as ankles and wrists) is less than that of larger joints (such as head and shoulders). To address the impact of scale variations across parts of the human body on the positioning of key points. In this paper, we propose a Detail Enhanced High-Resolution Network (DE-HRNet), which can efficiently extract local detail features and mitigate the impact of scale variations for human pose estimation. First, we propose a Detail Enhancement Module (DEM) to relearn the lost low-level detailed features and enhance the model's ability to capture delicate local features, which is crucial for improving the accuracy of scale-varying keypoints. Second, we introduce an ultra-lightweight dynamic sampler - dySample, which is used to replace nearest up-sampling. It aims to reduce the loss of detail information from low-resolution features during up-sampling, while simultaneously preserving finer local representations for high resolution, it can be beneficial in improving the robustness of the model in dealing with scale-varying keypoints. On the COCO test-dev2017 and MPII valid datasets, our method achieved 75.6 AP and 90.7 PCKh@0.5, respectively, compared to High-Resolution Network (HRNet), it improved by 0.7 and 0.4 points. In comparison with the other works, the proposed method has performed well in the scale variation.