Cascaded Pyramid Network (CPN)
This repo is also linked to tf-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 |
Usage
Train on MSCOCO dataset
- Clone the repository
git clone https://github.com/megvii-detection/tf-cpn
We'll call the directory that you cloned $CPN_ROOT.
-
Download MSCOCO images from http://cocodataset.org/#download. 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 config.py and you can modify them as you wish.
-
Download the base model (ResNet) weights from slim model_zoo and put them in $CPN_ROOT/data/imagenet_weights/.
-
Setup your environment by first running
pip3 install -r requirement.txt
- To train a CPN model, use network.py in the model folder.
python3 network.py -d 0-1
After the training finished, output is written underneath $CPN_ROOT/log/ which looks like below
log/
|->model_dump/
| |->snapshot_1.ckpt.data-00000-of-00001
| |->snapshot_1.ckpt.index
| |->snapshot_1.ckpt.meta
| |->...
|->train_logs.txt
Validation
Run the testing code in the model folder.
python3 mptest.py -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 mptest.py -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:
@article{Chen2018CPN,
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}
}
Third party implementation
Thanks for Geng David and his pytorch re-implementation of CPN.
Troubleshooting
- If you find it pending while running mptest.py, 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.
Contact
If you have any questions about this repo, please feel free to contact [email protected] or [email protected].