Introduction

Training Process.

Note for a bug, stage2 and stage 4's output are noise, which has already been fixed.

(click me to see the Training Process.)[http://oj5adp5xv.bkt.clouddn.com/trainng_process.png]

1. This is a mxnet-openpose implemention which is based on mxnet_Realtime_Multi-Person_Pose_Estimation by @dragonfly90.
   

2. I have tested this code under 4 K80 GPUS, after about 9000 iterations with batch 8, the loss can converge to about 90-110.
   

3. Pretrained models are in http://pan.baidu.com/s/1i5H2WHB

4. You can get more information from the original caffe version.

Prepare for train.

1. run mkdir model && mkdir models

2. Prepare a python2 environment, and install packages of matplotlib, scipy, mxnet(>0.9), numpy.

3. Go into cython folder, run python setup.py build_ext --inplace

4. Download MPI dataset from http://human-pose.mpi-inf.mpg.de(mpii_human_pose_v1.tar.gz,12.1GB,mpii_human_pose_v1_u12_2.zip), extracting mpii_human_pose_v1_u12_2.zip you'll get a matlab mat file named mpii_human_pose_v1_u12_1.mat. Extracting mpii_human_pose_v1.tar.gz you'll get a folder named imgaes.

5. Run python mpi_2_json.py --images=/data1/yks/dataset/openpose_dataset/mpi/images --annonations=/data1/yks/dataset/openpose_dataset/mpi/mpii_human_pose_v1_u12_2/mpii_human_pose_v1_u12_1.mat to convert mpii_human_pose_v1_u12_1.mat from format mat to a json file which will be putted into folder dataset,you may need to change the MPI images path and the path of the annonations file.

6. Run python mpi_parse.py to generate mask, heatmap and some other labels, this operation will generate a sqlite db file named mpi_inf_v2.db in folder dataset.

7. Run python train.py to train the model.
   

Demo

After you have trained your own model or download the pretrained model, you can run python2 demo.py --images=/data1/yks/dataset/openpose_dataset/mpi/images --prefix="models/yks_pose" --epoch=8600 to evaluate the model. ``