Domain Adaptation for anime face detection

This is an implementation of domain adaptation for anime face detection on Pytorch.

We referred hdjsjyl/face-faster-rcnn.pytorch.

Preparation

First of all, clone the code

git clone https://github.com/kanosawa/anime-face-faster-rcnn-da.pytorch.git

Then, create a folder:

cd anime-face-faster-rcnn-da.pytorch && mkdir data

Data Preparation

1. WIDER Face dataset
2. CelebA dataset : img_align_celeba.zip
3. animeface-character-dataset

Download above data and extract to data folder as below structure.

• data
  • WIDER2015
    • eval_tools
    • wider_face_split
    • WIDER_test
    • WIDER_train
    • WIDER_val
  • img_align_celeba
  • animeface-character-dataset

Pretationed Model

Download VGG16 and put them into the data/pretrained_model/

Compilation

cd lib
python setup.py build develop

Train

python trainval_net.py 

Demo

Put images to images folder and execute below command.

python demo.py --checksession $SESSION --checkepoch $EPOCH --checkpoint $POINT

if you want to use output/vgg16/wider_face/faster_rcnn_1_3_6439.pth, substitute $SESSION=1, $EPOCH=3, $POINT=6439