Realtime_multi-person_pose_estimation: will you open source the training code?

[gaopeng-eugene]

[ZheC]

Yes, we will release the training code in this repository soon (about two weeks later).

[sndler]

pretty cool

[sndler]

I am very interested in your methods, and how long will you open sources the training code? Thanks

[ShangxuanWu]

Is there any special layers in your training network? If not, I could rewrite it in tensorflow.

[power0341]

Awesome demo! and can't wait to try the training code. thanks

[sndler]

Thanks for your reply! there is no special layer. I have read your paper, but have some questions ,something not understand, so I want to understand your work better using your training code. My email address is [email protected] Sincerely thanks!

[ZheC]

Thank you all for the interest. We are trying our best to make training code easy for you guys to use or modify. It seems that we need a little more time to do this because we have tested hundreds of experiments, making our performance better but the code messier.

For those who cannot wait longer to try the training code, here is a temporary repo for our modified caffe: https://github.com/CMU-Perceptual-Computing-Lab/caffe_train I update the explanation about the training process in the README.

We will soon merge it with our testing caffe soon (https://github.com/CMU-Perceptual-Computing-Lab/caffe_rtpose). We will focus on making a clean training version for easier usage rather than maintaining the temporary repo (it will be replaced in the end). Thanks for understanding!

[sndler]

Much Thanks!

[sndler]

when python SetLayers.py, there is an error as following:
ERROR :AttributeError: 'LayerParameter' object has no attribute 'cpm_transform_param'
How to solve this error ,Thanks!

[ZheC]

In line 12 of setLayers.py, you need to specify the directory to caffe_train, instead of your own caffe. The attribute 'cpm_transform_param' is in our version of caffe.

[raymon-tian]

A new hand, can't wait to read the training code!

[sndler]

why the joints location is larger than the image's size ?
eg:
joint : #

but the frame's size is

Is the scale?

Thanks.

[gineshidalgo99]

The joints are scaled to the final frame resolution (by default 1280 x 720). You can modify the final resolution with the --resolution flag.

[ResByte]

Hi ZheC,
I am trying to train from scratch using your repo and caffe_rtpose on mscoco. I have been referencing your recent paper also. Can you share training details, as how long does it take to train(on what gpu)? what are the specific parameters(like input image size ) for training?

Thanks.

[ZheC]

Hi,

We have released the training parameters in https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation/tree/master/training

Specifically, you can set the parameters in https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation/blob/master/training/setLayers.py

pose_train_test.prototxt and pose_solver.prototx are example training files.

[shoCaffeTeria]

Great Work!
But I can't wait for your code to generate LMDB!!
Thanks

[ZheC]

@shoCaffeTeria I have added the code for generating LMDB in 'training' folder and the process in READEME file. I have tested them on my computer. Let me know if you meet any problem.

[shoCaffeTeria]

Thank u very much!
I want to try it soon, but my PC is in repairing now :-(

[ZheC]

@sndler I have fixed the issue of "AttributeError: 'LayerParameter' object has no attribute 'cpm_transform_param'" with an updated caffe_train(https://github.com/CMU-Perceptual-Computing-Lab/caffe_train). Also, it uses a newer version of caffe so that it is compatible with CUDA 7.5/ 8.0 and cudnn v5. Let me know if you meet other issues. Thanks!

[shoCaffeTeria]

Hi , ZheC
Your genLMBD code works well.
Thanks