| license: mit | |
| tags: | |
| - coreml | |
| - image-segmentation | |
| - cloth-segmentation | |
| # Cloth Segmentation — Core ML (U²-Net) | |
| On-device clothing segmentation for iOS, converted to Core ML. Segments a person's | |
| clothing into **upper-body**, **lower-body**, and **full-body (dress)** + background. | |
| - **Input:** RGB image, 768×768 (Core ML `ImageType`; normalization baked in). | |
| - **Output:** `probs` — softmax probabilities, shape `(1, 4, 768, 768)`, channels = `[background, upper, lower, full]`. | |
| - **Format:** FP16 mlprogram, ~84 MB. `ClothSegmentation.mlpackage.zip` (download + unzip + compile on device). | |
| ## Provenance & License | |
| - Architecture: **U²-Net** (Xuebin Qin et al.) — Apache-2.0. | |
| - Model/code: **levindabhi/cloth-segmentation** — MIT. | |
| - Weights trained on **iMaterialist (Fashion) 2019 @ FGVC6** (competition dataset). | |
| - Checkpoint mirror: `maiti/cloth-segmentation` (`cloth_segm_u2net_latest.pth`). | |
| - Converted with the included `convert.py` (PyTorch 2.2 → coremltools 8.3). | |
| Used by Chromaform (chromaform.art) for its Clothing Recolor tool. | |