Manga Face Clustering

Official PyTorch reimplementation of "Adaptation of Manga Face Representation for Accurate Clustering" presented in SIGGRAPH Asia 2018.

Environment

pip install -r requirements.txt

We conduct experiments using torch==1.8.0+cu111 and torchvision==0.9.0+cu111.

How to Use

Dataset Preparation

Download the Manga109 dataset from the official website. The annotation version that we used is v2018.05.31. You can see the version list here.

# script/split_dataset.py /path/to/Manga109_20xx_xx_xx
python script/crop_faces.py /path/to/Manga109_20xx_xx_xx
# script/split_dataset.py /path/to/Manga109_20xx_xx_xx

The commands that are commented out have already been executed.

The statistics of the test data can be obtained using the below command.

python script/get_data_stats.py /path/to/Manga109_20xx_xx_xx

Pre-training

If you want to pre-train a model by yourself, execute the following command.

python pre_train.py /path/to/Manga109_20xx_xx_xx
# evaluation of a pre-trained model
python eval.py /path/to/Manga109_20xx_xx_xx --model_path results/model.pth

If not, download our pre-trained model from here and put it to results/model.pth.

Fine-tuning

Set a value in [0, 10] as an argument of --title_idx.

python train.py /path/to/Manga109_20xx_xx_xx --title_idx 0

Performance

Method	Accuracy	NMI
Pre-train (paper)	0.48	0.63
Pre-train (reimpl.)	0.472	0.615
Fine-tune (paper)	0.64	0.71
Fine-tune (reimpl.)	0.666	0.718

Difference from the Original Implementation

• We implemented with PyTorch instead of Chainer. This is because ResNet-50 pre-trained on ImageNet for Chainer is not publicly available.
• The size of the dataset for pre-training (67,336) is somewhat larger than the value in our paper (67,328).
• The size of BEMADER_P (1,111 + 82) is somewhat larger than the value in our paper (1,105 + 82).

Links

• Our paper: https://dl.acm.org/citation.cfm?id=3283319
• manga109api: https://github.com/manga109/manga109api