Zhe Cao, Tomas Simon, Shih-En Wei, Yaser Sheikh.

batch_size=12
lr_schedule=mx.lr_scheduler.FactorScheduler(step*10,0.333)
if optim=='sgd':
    optimizer_params={ 'momentum':0.9,
                        'wd':  0.00005,
                        'learning_rate':lr,
                        'lr_scheduler':lr_schedule}
    print 'using sgd optimizer'
    print 'optimal prarams:',optimizer_params

i didn't caculate, the general is 5~6 second 5 batches with batch size 12. i think it is similar to the original code.

[x] loss during training

[x] some samples from the dataset

• first download the MPII dataset unzip it
  • some tips can refer https://github.com/kohillyang/mx-openpose
• if you use other dataset, please reimplement the MPII to ***, keep the roi_rec format unchanged. i adopt from VOC
• if you use other dataset like MSCOCO, the mask may changed, please change the dataiter file from line_90 to line_105
• the width of part affine field and sigma, and all other configration in the config.py.

mxnet_Realtime_Multi-Person_Pose_Estimation model_file ,and some others.

mx-openpose mask generation and some others.

Please cite the paper in your publications if it helps your research:

@inproceedings{cao2017realtime,
  author = {Zhe Cao and Tomas Simon and Shih-En Wei and Yaser Sheikh},
  booktitle = {CVPR},
  title = {Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields},
  year = {2017}
  }
  
@inproceedings{wei2016cpm,
  author = {Shih-En Wei and Varun Ramakrishna and Takeo Kanade and Yaser Sheikh},
  booktitle = {CVPR},
  title = {Convolutional pose machines},
  year = {2016}
  }