yolov11x-kaokore-face

YOLOv11x KaoKore Face

Model Description

A YOLOv11x detector fine-tuned to localize faces in Japanese artworks (浮世絵・肉筆画・近代美人画など). Photographic face detectors (Haar, RetinaFace, YuNet) miss the highly stylized faces of Japanese art (e.g. Utamaro's hikime-kagibana 引目鉤鼻). This model fills that gap by fine-tuning on face bounding boxes reconstructed from the 顔貌コレクション / KaoKore dataset (CODH).

• Single class: face
• Trained on KaoKore-reconstructed bounding boxes (≈9,683 face crops → detection dataset).

Intended Uses

• Face detection / localization in Japanese historical & early-modern artworks
• Building face-image collections (顔貌コレクション) from digitized art
• Pre-processing for downstream attribution/attribute classification (gender, social status, etc.)

How to Use

from ultralytics import YOLO

# Load model (downloads from the Hub)
model = YOLO("nakamura196/yolov11x-kaokore-face")

# Run inference
results = model.predict("your_image.jpg", conf=0.25, iou=0.45)

for result in results:
    print(result.boxes)

Download the weights directly:

from huggingface_hub import hf_hub_download

pt_path   = hf_hub_download(repo_id="nakamura196/yolov11x-kaokore-face", filename="best.pt")
onnx_path = hf_hub_download(repo_id="nakamura196/yolov11x-kaokore-face", filename="best.onnx")

The best.onnx export is suitable for ONNX Runtime (incl. onnxruntime-web in the browser).

best-fp16.onnx (114 MB) is the same model with fp16 weights (inputs and outputs stay float32), converted with onnxconverter-common for in-browser use. On a test print (NDL 1301975, 1280 px) its boxes match best.onnx within one pixel. It is the file used by the Inference tab of IIIF Kasane.

Training Data

Fine-tuned on a detection dataset reconstructed from the KaoKore dataset (顔貌コレクション, CODH) — face images (256×256, v1.3) cropped from Japanese historical artworks via IIIF, with gender (男/女) and social-status (貴族/武家/化身/庶民) labels. The KaoKore dataset is released under CC BY-SA 4.0.

Evaluation

Validation metrics on the held-out split (best epoch):

Model Architecture

• Base Model: YOLOv11x (Ultralytics, extra-large variant)
• Task: Object Detection (1 class: face)
• Framework: Ultralytics
• Formats: PyTorch (best.pt), ONNX (best.onnx)

Limitations

• Optimized for Japanese artwork; not intended for photographic face detection.
• Very small faces in dense group scenes may be missed — consider tiled inference (SAHI) for high-resolution images.
• Performance varies with style, period, and scan quality.

Acknowledgments & Dataset Attribution

This model is a derivative work of the KaoKore dataset, which is licensed under CC BY-SA 4.0 by the Center for Open Data in the Humanities (ROIS-DS CODH). The license requires attribution. The required credit is:

"KaoKore Dataset" (collected by CODH from multiple organizations), doi:10.20676/00000353

KaoKore itself is derived from the Collection of Facial Expressions (顔貌コレクション), CODH.

Citation

If you use this model, please cite the KaoKore dataset paper:

@inproceedings{tian2020kaokore,
  title     = {{KaoKore: A Pre-modern Japanese Art Facial Expression Dataset}},
  author    = {Yingtao Tian and Chikahiko Suzuki and Tarin Clanuwat and Mikel Bober-Irizar and Alex Lamb and Asanobu Kitamoto},
  booktitle = {Proceedings of the International Conference on Computational Creativity (ICCC)},
  year      = {2020},
  pages     = {415--422}
}

License

This model is released under AGPL-3.0, inherited from its base model Ultralytics YOLOv11. The training data (KaoKore) is licensed under CC BY-SA 4.0 by ROIS-DS CODH; per CC BY-SA, attribution (above) is required and derivative works are shared under the same license.