SCOP models
Model catalogue and mirrored weights for SCOP
(HEIG-VD, INSIT). SCOP reads models.json only when the user explicitly asks
for updates (scop.utils.models.check_updates(), or "Check for updates" in the
"Models and caches" dialog). It never updates a model silently.
models.json
Schema 1: {"schema": 1, "updated", "models": [{"id", "version", "contract", "url", "sha256", "size", "licence", "commercial_ok", "notes"?}]}. Every file
is checked against its SHA-256 after download. contract changes when the
inputs or outputs of a model change: a SCOP version only uses entries with
the contract it was written for.
Most entries point to their original source (Ultralytics-based detectors on
SWITCHdrive, LCD weights on GitHub, skyseg on Hugging Face). This repository
hosts the files that have no other public source, laid out as
<id>/<version>/<file>.
Hosted files
| File | Model | Licence |
|---|---|---|
sky.upernet_swin_t_ade/2026.10.31/upernet_swin_t_ade.onnx |
UperNet + Swin-T trained on ADE20K (openmmlab/upernet-swin-tiny), exported to ONNX at 576 px with the softmax in the graph; SCOP keeps the sky class |
MIT |
sky.upernet_swin_t_ade_fast/2026.10.31/upernet_swin_t_ade_fast.onnx |
The same network exported at 384 px (about 2x faster on CPU) | MIT |
The ADE20K training data is licensed for research and education. Check that your use is compatible before using these segmentations commercially.