基于密集关键点回归的人体姿态估计

Human Pose Estimation Based on Dense Keypoint Regression

针对人体姿态估计中直接回归坐标方法缺乏鲁棒性,精度较差等问题,总结了历年以来基于坐标回归的表征方法,提出了密集回归人体关节点的新型表征方法,增加关节点融合模块,同时兼顾速度与精度。在数据集COCO test-dev2017上获得了71.5%的得分,比HRNet基准模型提高了1%。

In human pose estimation,the direct regression of coordinates method lacks robustness and accuracy.The coordinate regression-based representation methods were summarized over the years and a novel representation method was proposed for densely regressing human keypoints,incorporating a joint fusion module while balancing speed and accuracy.On the COCO test-dev2017 dataset,a score of 71.5%was achieved,and improved by 1%over the HRNet benchmark model.