联合人体解析与姿态估计的轻量级网络

A Lightweight Network for Joint Human Parsing and Pose Estimation

人体解析与姿态估计是人类行为理解领域中两个重要的研究方向。其中,人体解析旨在区分人体图像的各个区域,而姿态估计的目标则是在图像中找出人的关节点。由于这两个任务存在天然的相关性,采用一个统一的模型同时实现两个任务,可以使两者相互受益并节省资源消耗。针对该问题,旨在设计一个高效的轻量级网络,以较少的计算资源实现两个任务一致性的高性能。在公开数据集LIP上的实验表明,提出的算法能加快推理速度并具备优良的性能。

Human parsing and pose estimation are two important research directions in the field of human behavior understanding.Among them,human body parsing aims to distinguish various regions of human images,and the goal of pose estimation is to find the joint points of people in the image.Due to the natural correlation between these two tasks,adopting a unified model to realize both tasks can benefit each other and save resource consumption.Aiming at this problem,this paper aims to design an efficient lightweight network that achieves high performance consistent with two tasks with less computing resources.Experiments on the public dataset LIP show that the algorithm proposed in this paper can speed up inference and have excellent performance.