Support multiple pose estimation

[pharrellyhy]

Hi @namedBen ,

Thanks for sharing this great repo. As far as I know, CPM only supports single pose estimation in the image. The follow-up paper of CPM, OpenPose, does support that but it is hard to implement and need lots of trick. Do you have any idea that we can extend current CPM to support two pose estimation? Thanks!

[namedBen]

First of all，open pose is very easy to implement.You can run the demo from https://github.com/tensorboy/pytorch_Realtime_Multi-Person_Pose_Estimation.(I tried and it worked well).I'm not sure the sentence "extend current CPM to support two pose estimation" you mean,maybe you can explain in more detail.～。～

[pharrellyhy]

Hi, @namedBen ,

Thanks for your quick reply. The network itself is not hard to implement but I got confused with the preprocessing of the training data since I want to train my own dataset, e.g., egocentric hands.

Basically what I want to achieve is to detect all the keypoints of all hands in the image so I need to train the model from scratch.

1. Can CPM localize keypoints of multiple objects, say person or hand?
2. If you familiar with how to train openpose from scratch, can you show me some hints, e.g., how to generate heatmap and paf? I also tried the demo of that repo but it does not provide how to generate the training data. That's the part that I confused most.

Thanks!

[namedBen]

Answer1：CPM can do it when you use dataset with multiple objects like COCO.(I tried before) Answer2: Actually, it's easy to generate heatmap and paf as long as you read the demo of PAFs carefully. At first, I tried to recurrent CPMs but failed.Then I read the demo of PAFs and implemented it successfully. Finally, I understand the two networks in turn.～。～ BTW,i can show you how to generate heatmap and paf but how can I contact you.....(you know,the Chinese Wall....)

[pharrellyhy]

Haha，can you send me your wechat ID？My email is pharrellyang90@gmail.com.

On Thu, Sep 20, 2018 at 8:04 AM namedBen notifications@github.com wrote:

Answer1：CPM can do it when you use dataset with multiple objects like COCO.(I tried before) Answer2: Actually, it's easy to generate heatmap and paf as long as you read the demo of PAFs carefully. At first, I tried to recurrent CPMs but failed.Then I read the demo of PAFs and implemented it successfully. Finally, I understand the two networks in turn.～。～ BTW,i can show you how to generate heatmap and paf but how can I contact you.....(you know,the Chinese Wall....)

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/namedBen/Convolutional-Pose-Machines-Pytorch/issues/3#issuecomment-422996670, or mute the thread https://github.com/notifications/unsubscribe-auth/AP5IsVbWukI-hmzgLZg3ySoeCTmvAh1jks5uctt2gaJpZM4Wvf_a .

--

Regards, Pharrell

[namedBen]

I have sent the email to u

[1209805200]

@namedBen 老板可以也发给我一份吗我最近在学习姿态估计相关嘿嘿，谢谢老板 1209805200@qq.com