AnimeCV

Pretrained computer vision tools for anime style illustrations. 日本語README

Getting started

Install AnimeCV by pip:

pip install git+https://github.com/kosuke1701/AnimeCV.git

Features

Near Human-Level Character Identification

The task is to identify which character is depicted in a picture given a set of reference pictures for each character. Example code

• Update on 2021.01.11
  • I released new pre-trained model for character face embedding!
  • See release and demo .
• Update on 2021.02.07
  • I released new pre-trained models for character face embedding.
  • The models are trained on ZACI-20, a public zero-shot character identification dataset.
  • The best model achieves near human-level performance with only 1.5 times larger error rate than that of humans.
  • You can load this model by using animecv.general.create_OML_ImageFolder_Encoder(model_directory_name).
  • Download models from release.

Object detection

The task is to detect objects in a picture and return bounding boxes around the objects.

• Face detector. Example code
  • Update on 2021.01.18
    • I released Yet Another Character Face Annotations on Danbooru2020 using this pre-trained model!
    • Automatically annotated face bounding boxes for the SFW 512px downscaled subset of Danbooru2020 dataset.

Note

All models are trained on datasets which mainly consists of Japanese anime style illustrations. Thus, it may not perform well on illustrations with other styles.