A Faster-RCNN based anime face detector.
This detector is trained on 6000 training samples and 641 testing samples, randomly selected from the dataset which is crawled from top 100 pixiv daily ranking.
Thanks to OpenCV based Anime face detector written by nagadomi, which helps labelling the data.
The original implementation of Faster-RCNN using Tensorflow can be found here
- Python 3.6 or 3.7
tensorflow
< 2.0opencv-python
cython
(optional, can be ignored with additional-nms-type PY_NMS
argument)- Pre-trained ResNet101 model
- Clone this repository
git clone https://github.com/qhgz2013/anime-face-detector.git
- Download the pre-trained model Google Drive: here Baidu Netdisk: here
- Unzip the model file into
model
directory - Install python & pipenv
pyenv install pipenv install
- Build the CPU NMS model (skip this step if use PY_NMS with argument:
-nms-type PY_NMS
)If using Windows Power Shell, typepipenv run setup
cmd /C make.bat
to run build script. - Run the demo as you want
- Visualize the result (without output path):
pipenv run main -i /path/to/image.jpg
- Save results to a json file
Format:
pipenv run main -i /path/to/image.jpg -o /path/to/output.json
{"image_path": [{"score": predicted_probability, "bbox": [min_x, min_y, max_x, max_y]}, ...], ...}
Sample output file:{"/path/to/image.jpg": [{"score": 0.9999708, "bbox": [551.3375, 314.50253, 729.2599, 485.25674]}]}
- Detecting a whole directory with recursion
pipenv run main -i /path/to/dir -o /path/to/output.json
- Customize threshold
pipenv run main -i /path/to/image.jpg -nms 0.3 -conf 0.8
- Customize model path
pipenv run main -i /path/to/image.jpg -model /path/to/model.ckpt
- Customize nms type (supports CPU_NMS and PY_NMS, not supports GPU_NMS because of the complicated build process for Windows platform)
pipenv run main -i /path/to/image.jpg -nms-type PY_NMS
- Visualize the result (without output path):
Mean AP for this model: 0.9086
Copyright info: 【C94】桜と刀 by 幻像黒兎
Copyright info: アイドルマスター シンデレラガールズ by 我美蘭@1日目 東A-40a
This model is directly trained by Faster-RCNN, with following argument:
python tools/trainval_net.py --weight data/imagenet_weights/res101.ckpt --imdb voc_2007_trainval --imdbval voc_2007_test --iters 60000 --cfg experiments/cfgs/res101.yml --net res101 --set ANCHOR_SCALES "[4,8,16,32]" ANCHOR_RATIOS "[1]" TRAIN.STEPSIZE "[50000]"
We've uploaded the dataset to Google drive here, dataset structure is similar to VOC2007 (used in original Faster-RCNN implementation).
Feel free to cite this repo and dataset. This work is not related to my research team and lab, just my personal interest.