YOLO_PPAP

PP/AP identification using YOLO_ver2 and chainer. For detail, see Qiita Entry.

Preparation

git clone this_repositry
ln -s /path/to/darknet .
ln -s /path/to/darknet/data/labels ./data/labels

Create materials

• Create your own PP/AP (PNG files with transparent background)
• Place them to data/ppap/foreground/00/ and data/ppap/foreground/01/
• Create your own background images
• Place them to data/ppap/background

Create dataset

cd data/ppap
mkdir images_pre
mkdir images
mkdir labels
python create_pretrain_dataset.py
python create_dataset.py
cd ../..

Pre-train & conert pre-train weights to initial weight

mkdir /tmp/backup
./pretrain.sh
./convert.sh

Train

mkdir /tmp/ppap-backup
./train.sh
cp /tmp/ppap-backup/tiny-yolo-final.weights ./YOLOtiny_chainer_v2

Convert darknet weights to chainer

cd ./YOLOtiny_chainer_v2
python YOLOtiny.py
cd ..

Prediction

predict image file

python replay_file.py filename.png

Output is written to filename_out.png

predict image files in data/ppap/images/

mkdir outfiles
python replay.py

Output is written to outfiles foldes

predict avi file

python replay_movie.py filename

Output is written to out.avi