garciafido/minifasnet-v2-anti-spoofing-onnx

MiniFASNet-V2 (ONNX)

ONNX export of the MiniFASNet-V2 (2.7_80x80) face anti-spoofing model from minivision-ai/Silent-Face-Anti-Spoofing.

What this is

A single-image face anti-spoofing classifier — given an 80×80 BGR face crop, returns a 3-class softmax [live, print-attack, replay-attack]. Used to gate identification flows against printed-photo / screen-replay attacks.

This repo redistributes the upstream weights converted to ONNX opset 11 so they can be loaded with onnxruntime (no PyTorch dependency at inference time). The model weights themselves are bit-equivalent to the upstream .pth — only the serialization format changed.

Provenance

• —Upstream weights: `2.7_80x80_MiniFASNetV2.pth` — SHA-256 a5eb02e1843f19b5386b953cc4c9f011c3f985d0ee2bb9819eea9a142099bec0.
• —Upstream architecture: MiniFASNetV2(embedding_size=128, conv6_kernel=(5,5), drop_p=0.2, num_classes=3, img_channel=3) per minivision-ai/Silent-Face-Anti-Spoofing/src/model_lib/MiniFASNet.py.
• —Conversion: torch 2.2.2 → ONNX opset 11, torch.onnx.export with dynamic batch axis. Conversion script: see convert_minifasnet_to_onnx.py reproduced from garciafido/agilface's scripts/ directory.
• —ONNX SHA-256: d7b3cd9ba8a7ceb13baa8c4720902e27ca3112eff52f926c08804af6b6eecc7b
• —Size: 1,744,116 bytes.

Preprocessing

Input shape: (1, 3, 80, 80), float32, BGR, range [0.0, 1.0] (i.e. pixel / 255).

The reference pipeline:

1. 1.Detect a face (e.g. with OpenCV's YuNet) and obtain the bounding box.
2. 2.Crop the image with a 2.7× scale margin around the bbox center (matches the upstream filename prefix 2.7_80x80).
3. 3.Resize the crop to 80×80, BGR, no alignment warp.
4. 4.Normalize: pixel / 255 → [0, 1].
5. 5.Reshape HWC → NCHW, run inference.
6. 6.Apply softmax over the 3-class output. Liveness score = 1 - (p[print] + p[replay]).

A reference implementation lives at `packages/infrastructure-py/src/agilface_infrastructure/ml/liveness_detector.py` in the AgilFace repo.

License

Apache 2.0 — same license as the upstream model. The original LICENSE file is included in this repo (LICENSE). Attribution: minivision-ai (Beijing Mininglamp Vision Technology Co., Ltd.). This repo does not claim authorship or training of the model — only the format conversion.

Citation

If you use this model, cite the upstream work:

@misc{minivisionai2020silentantispoofing,
  title  = {Silent-Face-Anti-Spoofing},
  author = {Mininglamp Vision Technology Co., Ltd.},
  year   = {2020},
  howpublished = {\\url{https://github.com/minivision-ai/Silent-Face-Anti-Spoofing}},
}