| pipeline_tag: image-segmentation | |
| tags: | |
| - onnx | |
| - roomorph | |
| - semantic-segmentation | |
| - browser | |
| # RooMorph room surface segmentation (ONNX) | |
| Browser model used by RooMorph to classify room pixels as other, wall, floor, or ceiling. This repository hosts the existing exported weights; it is not a new training run. | |
| ## Files and identity | |
| - `model.onnx`: 23,288,287 bytes. | |
| - SHA-256: `72d176cbe3e111551a04b21929a4c3bc564c24ed57c45f1529519323f6571ac7`. | |
| - `metadata.json`: input/output contract, preprocessing, class labels, and model identity. Use this file as the implementation reference. | |
| - `provenance.json`: retained source, training-data provenance declarations, and limitations. | |
| - `MODEL_IMPLEMENTATION_LICENSE.txt`: retained implementation license notice. | |
| ## Inference | |
| The model is a custom trained DeepLabV3+ MobileNet export. The float32 RGB input is NCHW `[1, 3, 513, 513]`, letterboxed as specified in metadata, scaled by 1/255 and normalized with ImageNet mean and standard deviation. RooMorph performs inference in the user's browser. This repository provides download hosting, not an inference API. | |
| ## Provenance and limitations | |
| The existing provenance identifies Qualcomm / AIMET model zoo pretrained weights and a Roboflow room segmentation dataset. Its declared licenses and the original implementation notice are retained without assigning a new license to the weights. The dataset provenance records that source-image rights and commercial rights were not independently verified. Consult `provenance.json` and the original sources for details. | |
| No new accuracy, benchmark, robustness, or commercial-rights claims are made by this hosting migration. Outputs can require correction and should be interpreted within RooMorph's visualization workflow. | |
| Production consumers should pin downloads to a full commit revision and verify the SHA-256 above. | |