Silica Gel Camera — gallery person model

Person segmentation model downloaded by the Silica Gel Camera iOS app the first time the user opens the photo library (gallery) with the "Ballad of You" filter. It is not bundled with the app to keep the app small.


File	`PersonRFDETRL504F16-v1.mlmodelc.aar`
Format	Compiled Core ML model (`.mlmodelc`) packed with Apple Archive (lzfse)
SHA-256	`fbfb2675fb4efe85d8bf1a6ab5aa3cce14fc0816fcabbc7699bfc6b33d507843`
Size	60,368,678 bytes
Input	`image`, RGB 0–255, 504×504
Output	`person`, GRAYSCALE_FLOAT16 person probability
Deployment target	iOS 17.0

The app checks the size and SHA-256 before installing; this file must not be replaced in place. A new model gets a new file name (-v2) and a matching app update.

Source

Model: RF-DETR-Seg Large (rf-detr-seg-large.pt, SHA-256 ca7b7c630ba22496067cc4f034c4e70c8e47fd7adcea04c3a49aa8c1755cbe6b) from rfdetr 1.11.2, © Roboflow, Inc., Apache-2.0.
Converted with coremltools 9.0 (float16). ImageNet normalization and the person-class softmax are inside the model; the head's 42→126 bilinear upsample is rewritten as two interpolation matrix products.

Build

From ios/ in the app repository:

uv run --python 3.12 --with torch==2.13.0 --with torchvision==0.28.0 --with transformers==5.18.0 \
    --with coremltools==9.0 --with pillow --with rfdetr==1.11.2 python Tools/Models/convert.py <out>
xcrun coremlcompiler compile <out>/PersonRFDETRL504F16.mlpackage /tmp/gallery --platform iOS --deployment-target 17.0
aa archive -D /tmp/gallery/PersonRFDETRL504F16.mlmodelc -o PersonRFDETRL504F16-v1.mlmodelc.aar -a lzfse \
    -exclude-field uid,gid,flg,mtm,btm,ctm,xat

Compiling twice gives a slightly different coremldata.bin (metadata order), so a rebuilt archive has a different SHA-256; the hash above is the one for the file in this repository.

License

Apache License 2.0 — see LICENSE and NOTICE.

Downloads last month: -; Downloads are not tracked for this model. How to track