inviol browser video privacy tool — model and software notices

Head model: Awiros/person_and_head_detection, RT-DETRv4-S, C-RADIOv4-SO400M-distilled.
Source: https://huggingface.co/Awiros/person_and_head_detection
Pinned revision: 823ca54bbf1446173f6906198a5e8b2a3e090890
Checkpoint: ckpts/rt-detrv4s-cradiov4-so400m/best_stg2.safetensors
Published license: Apache License 2.0. See privacy-head-LICENSE.txt.

Architecture: RT-DETRv4, https://github.com/RT-DETRs/RT-DETRv4
Pinned commit: 55fefaaed7efe2a5f72d0a18fd4e05965e35c292
Copyright (c) 2024 The DEIM Authors; Copyright (c) 2024 The D-FINE Authors;
Copyright (c) 2023 lyuwenyu. Apache License 2.0.
Modification by inviol: converted the published checkpoint to static-batch-1,
640x640 FP32 ONNX (opset 17), returning sigmoid scores and normalized boxes.
No retraining. The inference model does not include or run the teacher.
Export recipe: scripts/export-privacy-model.py in the website source.

ONNX Runtime Web 1.23.2: Microsoft Corporation, MIT License.
https://github.com/microsoft/onnxruntime
See onnxruntime-LICENSE.txt.

Mediabunny 1.21.0: Vanilagy, Mozilla Public License 2.0.
https://github.com/Vanilagy/mediabunny/tree/v1.21.0
See mediabunny-LICENSE.txt. Used unmodified from the published npm package.

This model detects regions of heads; it does not identify people.
All three presets use this same checkpoint. Fast: full frame. Normal:
full frame plus four overlapping crops. Max: full frame plus nine crops.
Published benchmark results are not a guarantee for a particular video.