摘要
人体姿态估计方法分为基于坐标回归的方法和基于热图的方法.基于坐标回归的方法推理速度较快但精度较差,基于热图的方法可精确定位,但计算量和存储开销较大.因此,文中通过知识蒸馏,结合两种方法,提出基于知识蒸馏与动态区域细化的人体姿态估计方法.首先,在特征蒸馏与姿态蒸馏两方面将热图模型的信息传递给回归模型.然后,对经过多层Transformer提取的特征进行选择,在粗略化阶段根据提取的特征生成初步姿态估计,并依据质量预测器的得分挑选需要细化的图像特征.最后,在细化阶段根据关键点与图像区域之间的相关程度,在部分关键点相关区域上建立细粒度表示,即细化特征,实现人体姿态细化.在COCO、COCO-Wholebody数据集上的实验表明,文中方法可较好地定位关键点,完成人体姿态估计.
Human pose estimation methods are categorized into coordinate regression-based methods and heatmap-based methods.Coordinate regression-based methods are characterized by slightly faster inference speed but slightly lower accuracy,while heatmap-based methods can achieve precise localization at the cost of higher computational and storage overhead.Therefore,a human pose estimation method based on knowledge distillation and dynamic region refinement is proposed.First,the information from the heatmap model is transferred to the regression model through feature distillation and pose distillation.Then,the features extracted by multi-layer Transformer are selected to generate initial pose estimation in the coarse stage,and the image features that need to be refined are selected based on the scores from a quality predictor.Finally,in the refinement stage,fine-grained representations or refined features,are established in the regions related to some keypoints according to the correlation between keypoints and image regions,achieving human pose refinement.Experiments on COCO and COCO-WholeBody datasets demonstrate that the proposed method can accurately locate keypoints and achieve accurate human pose estimation.
作者
魏龙生
付兴朋
李唐强
黄浩宇
WEI Longsheng;FU Xingpeng;LI Tangqiang;HUANG Haoyu(School of Automation,China University of Geosciences(Wuhan),Wuhan 430074)
出处
《模式识别与人工智能》
北大核心
2025年第2期164-176,共13页
Pattern Recognition and Artificial Intelligence
基金
国家自然科学基金项目(No.61603357)资助。
关键词
姿态估计
坐标回归
知识蒸馏
特征选择
特征细化
Pose Estimation
Coordinate Regression
Knowledge Distillation
Feature Selection
Feature Refinement
