Commercial-use and training-data clarification for SFace 2021dec ONNX weights

[jirayuu5119]

Hello OpenCV Zoo maintainers,

We are evaluating the following exact model for a locally hosted commercial photography-gallery face-similarity search:

• File: models/face_recognition_sface/face_recognition_sface_2021dec.onnx
• Git LFS SHA-256: 0ba9fbfa01b5270c96627c4ef784da859931e02f04419c829e83484087c34e79
• Size: 38,696,353 bytes
• Introducing commit: ba91a3b

The inference service would run only on hardware controlled by our photography business. It would rank event photos that may contain the same face as a customer-provided selfie. It would not be used for authentication or identity verification, and no external face API would receive customer data.

The model directory README states that all files are licensed under Apache-2.0. We also found:

• The original SFace repository mentions CASIA-WebFace, VGGFace2, and MS1MV2, but does not map this exact ONNX weight to a training dataset.
• In DNN model used in SFace feature extractor? #107, the original contributor confirmed that the OpenCV version is based on MobileNet/MobileFaceNet.
• The pull request that introduced the December 2021 weight does not document its training-data provenance.

Before commercial deployment, could an authorized maintainer or the original model contributor please clarify:

1. Does the Apache-2.0 license in this directory explicitly cover the pretrained weights in the exact ONNX file and checksum above, including commercial inference use?
2. Which exact training dataset or combination of datasets was used to produce this weight?
3. Do the applicable dataset/model terms permit commercial inference using the derived weight?
4. Is local use in a paid photography-gallery service permitted without a separate model license or fee?
5. May the weight be placed in a private application container or private model volume, provided that the Apache-2.0 license and attribution notices are preserved?
6. Is the Git LFS SHA-256 above the preferred immutable identifier for this model, or is there an official release/tag/checksum that should be used instead?

A public reply from an authorized maintainer or original contributor would be retained as part of our model compliance record.

cc @fengyuentau @zhongyy

Thank you for your time and for maintaining OpenCV Zoo.