AimmiA on-device ear-region segmentation

BiSeNet / ResNet18 face parsing, Copyright (c) 2024 Yakhyokhuja Valikhujaev.
https://github.com/yakhyo/face-parsing
ONNX weights mirrored at https://huggingface.co/jbrownkramer/face-parsing
Revision: 4be031c61a22e801ab389ab9ccf954781772fd42
Model and upstream code published under MIT; see LICENSE.txt.
The vendored model has been prepared by tools/fetch_tryon_ear_assets.py:
training auxiliary outputs removed, spatial dimensions made dynamic, and four
exported fixed-size global pools restored to GlobalAveragePool. Learned weights
are unchanged. Photo inference uses 512px; live inference uses 384px crops.

ONNX Runtime Web 1.20.1, Copyright (c) Microsoft Corporation.
https://github.com/microsoft/onnxruntime/tree/v1.20.1
MIT; see LICENSE.txt.

The model segments visible ear regions (classes 7 and 8 of CelebAMask-HQ).
The lower contour supplies an estimated lobe attachment point, not a detected
piercing hole. Occluded or unconfident ears are not used for placement.
Inference runs in a browser worker. Images are not uploaded.

Rebuild the vendored assets with: python3 tools/fetch_tryon_ear_assets.py