人体姿态估计两种主要思路

客官°小女子只卖身不卖艺 2024-04-17 23:16 152阅读 0赞

先挂上大佬（旷世余刚）的文章链接：[人体姿态估计的过去、现在和未来][Link 1] ，这篇把人体姿态估计介绍的很好很细了，本文主要对相关的一些代码做了一些链接引导，仅供参考。  
      人体姿态估计（Human Pose Estimation）也称为人体关键点检测（Human Keypoints Detection）。对于人体姿态估计的研究，主要有两种思路：Top-down 和Bottom-up

## Top-down方法 ##

Top-down首先利用目标检测算法检测出单个人，得到边界框，然后在每一个边界框中检测人体关键点，连接成一个人形，缺点就是受检测框的影响太大，漏检，误检，IOU大小等都会对结果有影响，算法包括RMPE、Mask-RCNN 等，这种方法一般具有较高的准确率但是处理速度较低，代表性的方法有：

**(1) Stacked Hourglass Network（半壁江山）**

Github: [https://github.com/Naman-ntc/Pytorch-Human-Pose-Estimation][https_github.com_Naman-ntc_Pytorch-Human-Pose-Estimation] (Pytorch)

[https://github.com/wbenbihi/hourglasstensorlfow][https_github.com_wbenbihi_hourglasstensorlfow] (Tensorflow)

**(2) Convolutional Pose Machine**

Github: [https://github.com/timctho/convolutional-pose-machines-tensorflow][https_github.com_timctho_convolutional-pose-machines-tensorflow] (Tensorflow)

[https://github.com/JindongJiang/WeightedLoss\_Convolutional\_Pose\_Machines\_PyTorch][https_github.com_JindongJiang_WeightedLoss_Convolutional_Pose_Machines_PyTorch] (Pytroch)

**(3) Cascaded Pyramid Network**

Github: [https://github.com/chenyilun95/tf-cpn][https_github.com_chenyilun95_tf-cpn]

**(4) Simple Baseline (proposed by MSRA)**

Github: [https://github.com/Microsoft/human-pose-estimation.pytorch][https_github.com_Microsoft_human-pose-estimation.pytorch]

**(5) Multi-Stage Pose Network (MSPN，由Face++提出，2018年coco 关键点检测冠军)**

--------------------

## Bottom-up方法 ##

首先对整个图片进行每个人体关键点部件的检测，再将检测到的部件拼接成一个人形，这种方法一般准确率较差，会将不同人的不同部位按一个人进行拼接，但处理速度较快，代表方法就是OpenPose、DeepCut 、PAF。

**(1) HighResolution Net (HRNet，CVPR 2019/Proposed by MSRA&USTC)**

Github: [https://github.com/leoxiaobin/deep-high-resolution-net.pytorch][https_github.com_leoxiaobin_deep-high-resolution-net.pytorch]

**(2) CMU的Realtime Multi-Person Pose Estimation(OpenPose)**

Github：[https://github.com/ZheC/Realtime\_Multi-Person\_Pose\_Estimation][https_github.com_ZheC_Realtime_Multi-Person_Pose_Estimation]

**(3) Newell（Hourglass的作者）的Associative embedding: End-to-end learning for joint detection and grouping**

Github：[https://github.com/princeton-vl/pose-ae-train][https_github.com_princeton-vl_pose-ae-train]

--------------------

**其他参考：**

*  常用数据集和主流方法介绍链接：[https://blog.csdn.net/buling5143/article/details/79466588][https_blog.csdn.net_buling5143_article_details_79466588]
 *  人体姿态估计(Human Pose Estimation)文献综述：[https://blog.csdn.net/BockSong/article/details/81037059][https_blog.csdn.net_BockSong_article_details_81037059]

[Link 1]: https://blog.csdn.net/dQCFKyQDXYm3F8rB0/article/details/102479871
[https_github.com_Naman-ntc_Pytorch-Human-Pose-Estimation]: https://github.com/Naman-ntc/Pytorch-Human-Pose-Estimation
[https_github.com_wbenbihi_hourglasstensorlfow]: https://github.com/wbenbihi/hourglasstensorlfow
[https_github.com_timctho_convolutional-pose-machines-tensorflow]: https://github.com/timctho/convolutional-pose-machines-tensorflow
[https_github.com_JindongJiang_WeightedLoss_Convolutional_Pose_Machines_PyTorch]: https://github.com/JindongJiang/WeightedLoss_Convolutional_Pose_Machines_PyTorch
[https_github.com_chenyilun95_tf-cpn]: https://github.com/chenyilun95/tf-cpn
[https_github.com_Microsoft_human-pose-estimation.pytorch]: https://github.com/Microsoft/human-pose-estimation.pytorch
[https_github.com_leoxiaobin_deep-high-resolution-net.pytorch]: https://github.com/leoxiaobin/deep-high-resolution-net.pytorch
[https_github.com_ZheC_Realtime_Multi-Person_Pose_Estimation]: https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation
[https_github.com_princeton-vl_pose-ae-train]: https://github.com/princeton-vl/pose-ae-train
[https_blog.csdn.net_buling5143_article_details_79466588]: https://blog.csdn.net/buling5143/article/details/79466588
[https_blog.csdn.net_BockSong_article_details_81037059]: https://blog.csdn.net/BockSong/article/details/81037059