Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time

Cascaded Pyramid Network (CPN)

This repo is also linked to megvii-cpn

This is a Tensorflow re-implementation of CPN (Cascaded Pyramid Network), which wins 2017 COCO Keypoints Challenge. The original repo is based on the inner deep learning framework (MegBrain) in Megvii Inc.

Results on COCO minival dataset (Single Model)

Note that our testing code is based on some detectors. In COCO minival dataset, the used detector here achieves an AP of 41.1 whose human AP is 55.3 in COCO minival dataset.

Method Base Model Input Size AP @0.5:0.95 AP @0.5 AP @0.75 AP medium AP large
CPN ResNet-50 256x192 69.7 88.3 77.0 66.2 76.1
CPN ResNet-50 384x288 72.3 89.1 78.8 68.4 79.1
CPN ResNet-101 384x288 72.9 89.2 79.4 69.1 79.9

Results on COCO test-dev dataset (Single Model)

Here we use the strong detector that achieves an AP of 44.5 whose human AP is 57.2 in COCO test-dev dataset.

Method AP @0.5:0.95 AP @0.5 AP @0.75 AP medium AP large
Detectron(Mask R-CNN) 67.0 88.0 73.1 62.2 75.6
CPN(ResNet-101, 384x288) 72.0 90.4 79.5 68.3 78.6

For reference, by using the detection results given by MegDet that achieves an AP of 52.1 whose human AP is 62.9, human pose result is as followed.

Method AP @0.5:0.95 AP @0.5 AP @0.75 AP medium AP large
MegDet+CPN(ResNet-101, 384x288) 73.0 91.8 80.8 69.1 78.7


Train on MSCOCO dataset

  1. Clone the repository
git clone

We'll call the directory that you cloned $CPN_ROOT.

  1. Download MSCOCO images from We train in COCO trainvalminusminival dataset and validate in minival dataset. Then put the data and evaluation PythonAPI in $CPN_ROOT/data/COCO/MSCOCO. All paths are defined in and you can modify them as you wish.

  2. Download the base model (ResNet) weights from slim model_zoo and put them in $CPN_ROOT/data/imagenet_weights/.

  3. Setup your environment by first running

pip3 install -r requirement.txt
  1. To train a CPN model, use in the model folder.
python3 -d 0-1

After the training finished, output is written underneath $CPN_ROOT/log/ which looks like below

       |    |->
       |    |->snapshot_1.ckpt.index
       |    |->snapshot_1.ckpt.meta
       |    |->...


Run the testing code in the model folder.

python3 -d 0-1 -r 350

This assumes there is an models that has been trained for 350 epochs. If you just want to specify a pre-trained model path, it's fine to run

python3 -d 0-1 -m log/model_dump/snapshot_350.ckpt

Here we provide the human detection boxes results:

Person detection results in COCO Minival

Person detection results in COCO test-dev

Pre-trained models:




Citing CPN

If you find CPN useful in your research, please consider citing:

    Author = {Chen, Yilun and Wang, Zhicheng and Peng, Yuxiang and Zhang, Zhiqiang and Yu, Gang and Sun, Jian},
    Title = {{Cascaded Pyramid Network for Multi-Person Pose Estimation}},
    Conference = {CVPR},
    Year = {2018}

You may also be interested in the following papers:


     title={Rethinking on Multi-Stage Networks for Human Pose Estimation},
     author={Li, Wenbo and Wang, Zhicheng and Yin, Binyi and Peng, Qixiang and Du, Yuming and Xiao, Tianzi and Yu, Gang and Lu, Hongtao and Wei, Yichen and Sun, Jian},
     journal={arXiv preprint arXiv:1901.00148},


       title={Learning Delicate Local Representations for Multi-Person Pose Estimation},
       author={Yuanhao Cai and Zhicheng Wang and Zhengxiong Luo and Binyi Yin and Angang Du and Haoqian Wang and Xinyu Zhou and Erjin Zhou and Xiangyu Zhang and Jian Sun},

Third party implementation

Thanks for Geng David and his pytorch re-implementation of CPN.


  1. If you find it pending while running, it may be the blocking problem of python queue in multiprocessing. For convenience, I simply implemented data transferring via temporary files. You need to call MultiProc with extra parameter "dump_method=1" and it'll be fine to run the test code with multiprocess.


If you have any questions about this repo, please feel free to contact


Cascaded Pyramid Network for Multi-Person Pose Estimation (CVPR 2018)







No releases published


No packages published