Skip to content

Latest commit

 

History

History

mindspore

README.md

Openpose

Model description

Openpose network proposes a bottom-up human attitude estimation algorithm using Part Affinity Fields (PAFs). Instead of a top-down algorithm: Detect people first and then return key-points and skeleton. The advantage of openpose is that the computing time does not increase significantly as the number of people in the image increases.However,the top-down algorithm is based on the detection result, and the runtimes grow linearly with the number of people.

Paper: Zhe Cao,Tomas Simon,Shih-En Wei,Yaser Sheikh,"Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields",The IEEE Conference on Computer Vision and Pattern Recongnition(CVPR),2017

Step 1:Installation

# Pip the requirements
pip3 install -r requirements.txt
wget https://download.open-mpi.org/release/open-mpi/v4.0/openmpi-4.0.7.tar.gz
tar xf openmpi-4.0.7.tar.gz
cd openmpi-4.0.7/
./configure --prefix=/usr/local/bin --with-orte
make -j4 && make install
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib/

Step 2:Preparing datasets

  • Go to visit COCO official website, then select the COCO dataset you want to download. Take coco2017 dataset as an example, specify /path/to/coco2017 to your COCO path in later training process, the unzipped dataset path structure sholud look like:

    coco2017
├── annotations
│   ├── instances_train2017.json
│   ├── instances_val2017.json
│   └── ...
├── train2017
│   ├── 000000000009.jpg
│   ├── 000000000025.jpg
│   └── ...
├── val2017
│   ├── 000000000139.jpg
│   ├── 000000000285.jpg
│   └── ...
├── train2017.txt
├── val2017.txt
└── ...

  • Create the mask dataset. Run python gen_ignore_mask.py

    python3 ./src/gen_ignore_mask.py --train_ann ./coco2017/annotations/person_keypoints_train2017.json --val_ann ./coco2017/annotations/person_keypoints_val2017.json --train_dir ./coco2017/train2017 --val_dir ./coco2017/val2017

  • The dataset folder is generated in the root directory and contains the following files:

    ├── coco2017
    ├── annotations
        ├─ person_keypoints_train2017.json
        └─ person_keypoints_val2017.json
    ├─ ignore_mask_train
    ├─ ignore_mask_val
    ├─ train2017
    ├─ val2017
    └─ ...

  • Download the VGG19 model of the MindSpore version:

    vgg19-0-97_5004.ckpt

Step 3:Training

Change the absolute path of the data in running shell train_openpose_coco2017_1card.sh train_openpose_coco2017_8card.sh.

For example in train_openpose_coco2017_1card.sh:

bash scripts/run_standalone_train.sh /home/coco2017/train2017 /home/coco2017/annotations/person_keypoints_train2017.json /home/coco2017/ignore_mask_train /home/vgg19-0-97_5004.ckpt

# Run on 1 GPU
bash train_openpose_coco2017_1card.sh

# Run on 8 GPU 
bash train_openpose_coco2017_8card.sh

# Run eval
python3 eval.py --model_path /home/openpose_train_8gpu_ckpt/0-80_663.ckpt --imgpath_val coco2017/val2017 --ann coco2017/annotations/person_keypoints_val2017.json

Results

GPUS AP AP  .5 AR AR  .5
BI V100×8 0.3979 0.6654 0.4435 0.6889

Reference

Openpose