Learning Delicate Pixel-Level Representations for Bottom-Up Human Pose Estimation
Xiaonan Wu, Zengzhao Chen, Hai Liu · 2021
The Gaussian heatmap regression, which is the mainstream technique, has already occupied a pivotal position in locating keypoints of the human body appearing in images or videos. Many researchers have conducted thorough and comprehensive studies, but we find that they rarely devote their mind to the impact of different pixel types on keypoint detection. Inspired by this discovery, this paper mainly studies the process of Gaussian heatmap construction and adopt a new technique in the bottom-up method to mitigate the issue of pixel imbalance. Additionally, AF1-measure, which is considered false positives and false positives in the evaluation process to balance AP and AR, is proposed to deal with scenarios where there is a huge gap between AP and AR during the experiment and researchers cannot evaluate the performance of the network model more accurately. Large experimental results show that our technique outstrips the advanced approaches in various evaluation metrics.