Cascaded Pyramid Network (CPN) based on keras (tensorflow backend)

Results on COCO minival dataset (Single Model)

Note that our testing code is based on detection results from original tf-cpn (In COCO validation, detector AP is 41.1 whose human AP is 55.3).

You can download our model here: https://github.com/Longqi-S/keras_cpn/releases/download/v0.1/cpn_resnet50_cpn_0065.h5

Prepare

1. 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.

   We use the human detection results same with tf-cpn, you can download it here: https://github.com/Longqi-S/keras_cpn/releases/download/v0.1/person_detection_minival411_human553.json.coco

   After preparation, file stucture should be like below:

data/
       |->COCO/
       |    |->dets/
       |    |    |->person_detection_minival411_human553.json.coco
       |    |->MSCOCO/
       |    |    |->PythonAPI/
       |    |    |->train2014/
       |    |    |->val2014/
       |    |    |->person_keypoints_minival2014.json
       |    |    |->person_keypoints_trainvalminusminival2014.json

2. Download the base model (ResNet) weights from [keras model_zoo]

cd $CPN_ROOT/data
sh get_pretrain_model.sh

3. Setup your environment by first running

pip3 install -r requirement.txt

Train

To train a CPN model, use train.py in root folder.

python3 train.py --model data/pretrain/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5 --cfg configs/e2e_CPN_ResNet50_FPN_cfg.py

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

logs/
       |->resnet50_cpn20180819T1607/
       |    |->events.out.tfevents.1534666165.9507
       |    |->cpn_resnet50_cpn_0001.h5
       |    |->cpn_resnet50_cpn_0002.h5
       |    |->...

Test

Run the testing code in the postprocessing folder.

cd postprocessing
python3 mptest.py -d 0 -m cpn_resnet50_cpn_0002.h5 -c ../configs/e2e_CPN_ResNet50_FPN_cfg.py

notice

We can only use just one GPU to test.

How to draw network architecture

Go to lib/utils/

python3 draw_net.py --mode 0 --cfg configs/e2e_CPN_ResNet50_FPN_cfg.py

mode: 0 means train; 1 means inference; cfg : choose which network to draw;