# Third-party components

Wifi Atlas is released under the [MIT licence](LICENSE). Everything it bundles is also open source (MIT or Apache-2.0),
so the app can be used, copied, modified and redistributed freely. The AI components are bundled with the app (same
origin) so that it has no runtime dependency on third-party servers. Full licence texts are in [`licenses/`](licenses/).

| Component | Used for | Size | Licence |
|---|---|---|---|
| [ONNX Runtime Web](https://github.com/microsoft/onnxruntime) 1.30.0 | Runs the models in the browser (WebGPU / WebAssembly) | ~27 MB | MIT, (c) Microsoft |
| [MoGe-2 ViT-S](https://github.com/microsoft/MoGe) (ONNX export from `Ruicheng/moge-2-vits-normal-onnx`) | Metric 3D geometry and surface normals from one camera image | 141 MB | MIT, (c) Microsoft; contains DINOv2 parts under Apache-2.0 |
| [Depth Anything V2 Small](https://github.com/DepthAnything/Depth-Anything-V2) (8-bit ONNX from `onnx-community/depth-anything-v2-small`) | Fallback depth model for phones without WebGPU | 27 MB | Apache-2.0 |

No semantic-segmentation model is used: rooms, walls, floors, ceilings and doors are derived from MoGe-2's geometry and
surface normals plus the phone's orientation sensors (`js/depthgeo.js`, `js/roomfit.js`). Models trained on ADE20K or
released under non-commercial licences (for example NVIDIA SegFormer) were deliberately left out to keep the whole stack
unambiguously free to use.

The Wi-Fi quality probe, the 3D viewer and everything else in this repository are original to this project (MIT).