Face detection for the 2D crowd

What it does This package detects near-frontal anime faces in images and predicts 28 facial keypoints—eyes, nose, mouth, and jawline—using pretrained models built atop OpenMMLab’s mmdetection and mmpose frameworks. It returns bounding boxes with confidence scores and dense landmark arrays for each detected face.

The interesting bit Most face-detection models are trained on human photographs and fail comically on illustrated characters. Rather than stop at a simple detector, the author used k-means clustering on the extracted 28-point landmarks to visualize mean face shapes across real images, turning a narrow utility into a small piece of domain analysis.

Key highlights

• Detects near-frontal anime faces and outputs 28 keypoints per face.
• Pretrained models download automatically on first use.
• Ships with a Gradio demo and a ready-to-run Colab notebook.
• Includes an exploratory k-means clustering of detected landmarks and the resulting mean face images.

Caveats

• Explicitly tested only on Ubuntu; portability elsewhere is unclear.
• Near-frontal poses only, so profiles and extreme angles are out of scope.
• Effectively a specialized wrapper around the full mmdetection and mmpose stack, so expect their heft and dependency complexity.

Verdict Worth grabbing if you are building anime-specific pipelines like auto-cropping avatars or analyzing illustration datasets. Skip it if you need a general-purpose human detector or a lightweight, dependency-free script.