面向姿态估计的组件感知自适应算法

Component-aware Adaptive Algorithm for Pose Estimation

针对姿态估计的结构化输出特点,提出一种领域自适应学习算法。建立一种组件感知的参数学习过程,根据目标的各组件调整自适应参数,提升模型的泛化能力。依据领域自适应算法特点引入基于主动学习样本选取策略,提升模型的学习效率。对特征分布差异的2种训练场景进行模拟实验,结果表明,该算法训练的模型在平均准确率上比传统学习算法提升6%～8%,比已有的自适应算法提升2%～4%,使用样本选取策略后,则进一步提升约2%。

This paper introduces a new domain adaptation algorithm for training pose estimation parameters that are characterized by structured output prediction.The algorithm establishes a component-aware parameter learning process,which can adjust the adaptive parameters according to the components of the target,and is suitable for the scenes with different feature distributions to improve the generalization ability of the pose estimation model.In order to further improve the learning efficiency of the model,an active learning sample selection strategy is introduced according to the characteristics of the domain adaptive algorithm.The experiments simulate two kinds of training scenarios with different feature distributions,i.e.,different pose distribution and different appearance distribution.Experimental results show that the average precision of the proposed algorithm is improved by 6%~8%compared with the traditional learning algorithms and 2%~4%compared with the existing adaptive algorithms,after using the sample selection strategies,it will increase by about 2%.