Training data / provenance of the res10 SSD face detector

,
1

Hi all,

We use OpenCV 4.5.3 (DNN module) with the bundled ResNet-10 SSD face detector — res10_300x300_ssd_iter_140000_fp16.caffemodel + deploy.prototxt from samples/dnn/face_detector. We run it for inference only (no training/fine-tuning on our side).

For EU AI Act documentation, as a deployer of this pre-trained model we need the provider’s info on how it was produced. Could anyone point us to:

  1. Training dataset — source, size, license (model card / datasheet)?
  2. Representativeness & bias — any known performance variation across skin tone, age, lighting, or pose?
  3. Evaluation metrics — published accuracy (e.g. WIDER FACE)?
  4. License of the model weights?

Links to official docs, the training repo, or the maintainer who produced it would be a big help. Thanks!

EU AI Act compliance