The dataset can be found here: https://drive.google.com/file/d/1HG7YnakUkjaxtNMclbl2t5sJwGLcHYsI/view?usp=sharing
This dataset should only be used for educational purposes.
Mckinsey666
This is an dataset consisting of 63632 high-quality anime faces scraped from www.getchu.com, which are then cropped using the anime face detection algorithm in https://github.com/nagadomi/lbpcascade_animeface. Images sizes vary from 90 * 90 ~ 120 * 120 (you can simply rescale them before using them).
Compared to other widely used datasets (such as the danbooru dataset, which is actually quite a mess), this dataset contains high quality anime character images with clean background and rich colors.
However, few outliers are still present in the dataset:
- Bad cropping results
- Some non-human faces.
Feel free to contribute to this dataset by adding images of similar quality or adding image labels.
I made the web-scraping and face-detection code open-source. Make modifications as you wish. To scrape the pictures directly from the website and detect the faces on your own, type the following commands:
git clone https://github.com/Mckinsey666/Anime-Face-Dataset.git
cd Anime-Face-Dataset/src
python3 scrape.py
python3 detect.py
The scraped images will be in the /src/images
folder, while the cropped images (character faces) will be in the /src/cropped
folder.