ajx91's picture
add core ml build
98ab458 verified
|
Raw History Blame Contribute Delete
1.02 kB
---
license: mit
base_model: ZhengPeng7/BiRefNet_HR-matting
pipeline_tag: image-segmentation
library_name: coreml
tags:
- coreml
- background-removal
- matting
---
# BiRefNet HR-matting for Core ML
[BiRefNet_HR-matting](https://huggingface.co/ZhengPeng7/BiRefNet_HR-matting) by Zheng Peng et al., converted to a Core ML ML Program (fp16) for macOS and iOS.
- `BiRefNetHRMatting-1.aar`: an Apple Archive of `BiRefNetHRMatting.mlpackage`. Extract with `aa extract -i BiRefNetHRMatting-1.aar -d .`.
- Input `image`: RGB, 2048 × 2048. ImageNet normalisation is inside the model.
- Output `alpha`: `[1, 1, 2048, 2048]` soft matte, 0 to 1.
- Run on `.cpuAndGPU`. The Neural Engine can't compile it.
- Deformable convolution is reimplemented with `grid_sample`, matching torchvision to about 1e-5. Mean matte difference from PyTorch is about 0.0003.
MIT licensed, like the original. Please cite the BiRefNet paper: Zheng et al., "Bilateral Reference for High-Resolution Dichotomous Image Segmentation", CAAI AIR 2024.