This is an anime face detector with facial landmark estimation, implemented in plain PyTorch. The face detectors (Faster R-CNN and YOLOv3) and the landmark model (HRNetV2, 28 points) were originally trained with mmdetection and mmpose; since v0.1.0 this package runs them without any OpenMMLab runtime dependency.
(To avoid copyright issues, the above demo uses images generated by the
TADNE model.)
The model detects near-frontal anime faces and predicts 28 landmark points.
The result of k-means clustering of landmarks detected in real images:
The mean images of real images belonging to each cluster:
pip install anime-face-detectorPython 3.12 or later is required. PyTorch (2.2.0 or later) and torchvision are installed as dependencies; if you need a specific CUDA build, install them first by following the official instructions.
import cv2
from anime_face_detector import create_detector
detector = create_detector("yolov3")
image = cv2.imread("assets/input.jpg")
preds = detector(image)
print(preds[0]){'bbox': array([2.2450244e+03, 1.5940223e+03, 2.4116030e+03, 1.7458063e+03,
9.9987185e-01], dtype=float32),
'keypoints': array([[2.2593938e+03, 1.6680436e+03, 9.3236601e-01],
[2.2825300e+03, 1.7051841e+03, 8.7208068e-01],
[2.3412151e+03, 1.7281011e+03, 1.0052248e+00],
[2.3941377e+03, 1.6825046e+03, 5.9705663e-01],
[2.4039426e+03, 1.6541921e+03, 8.7139702e-01],
[2.2625220e+03, 1.6330233e+03, 9.7608268e-01],
[2.2804077e+03, 1.6408495e+03, 1.0021354e+00],
[2.2969380e+03, 1.6494972e+03, 9.7812974e-01],
[2.3357908e+03, 1.6453258e+03, 9.8418534e-01],
[2.3475276e+03, 1.6355408e+03, 9.5060223e-01],
[2.3612463e+03, 1.6262626e+03, 9.0553057e-01],
[2.2682278e+03, 1.6631940e+03, 9.5465249e-01],
[2.2814783e+03, 1.6616484e+03, 9.0782022e-01],
[2.2987590e+03, 1.6692812e+03, 9.0256405e-01],
[2.2833625e+03, 1.6879142e+03, 8.0303693e-01],
[2.2934949e+03, 1.6909009e+03, 8.9718056e-01],
[2.3021218e+03, 1.6863715e+03, 9.3882143e-01],
[2.3471826e+03, 1.6636573e+03, 9.5727938e-01],
[2.3677822e+03, 1.6540554e+03, 9.4890594e-01],
[2.3889211e+03, 1.6611255e+03, 9.5125675e-01],
[2.3575544e+03, 1.6800433e+03, 8.5919142e-01],
[2.3688926e+03, 1.6800665e+03, 8.3275074e-01],
[2.3804905e+03, 1.6761322e+03, 8.4160626e-01],
[2.3165366e+03, 1.6947096e+03, 9.1840971e-01],
[2.3282458e+03, 1.7104808e+03, 8.8045174e-01],
[2.3380054e+03, 1.7114034e+03, 8.8357794e-01],
[2.3485500e+03, 1.7080273e+03, 8.6284375e-01],
[2.3378748e+03, 1.7118135e+03, 9.7880816e-01]], dtype=float32)}
The pretrained weights are hosted on the Hugging Face Hub as safetensors files (faster-rcnn, yolov3, hrnetv2) and are downloaded automatically on first use. The original .pth checkpoints from the mmdetection/mmpose days remain available in the v0.0.1 release. The anime-face-detector-convert command converts such checkpoints, including self-trained ones, to safetensors.
Demo (using Gradio)
pip install anime-face-detector gradio
git clone https://github.com/hysts/anime-face-detector
cd anime-face-detector
python demo_gradio.pyThe code in this repository is released under the MIT License (see LICENSE). The code under anime_face_detector/_vendor/ is derived from mmpose v0.29.0, mmdetection v2.28.2 and mmcv v1.7.0, which are licensed under the Apache License 2.0; see THIRD_PARTY_LICENSES.md for details.
If you find this repo useful for your research, please consider citing it:
@misc{anime-face-detector,
author = {hysts},
title = {Anime Face Detector},
year = {2021},
howpublished = {\url{https://github.com/hysts/anime-face-detector}}
}