duke182/tryon-garment-detector

Try-on garment detector (D-FINE-n, 3 classes)

Small garment detector for a browser try-on extension. Runs with ONNX Runtime Web on WebGPU (about 18 ms per image on an Apple M1).

Model contract

• Input pixel_values [1,3,320,320], RGB in 0..1, no mean/std. Letterbox to the top-left with a black pad.
• Output logits [1,100,3]: apply sigmoid (no "no object" class). Classes: 0 upper, 1 lower, 2 dress.
• Output pred_boxes [1,100,4]: cx, cy, w, h, normalized to the letterboxed input.
• Suggested score threshold: 0.7.

Use fp32 on WebGPU. An fp16 conversion gave wrong output on WebGPU in testing.

Evaluation (small, indicative only)

50 positive + 25 negative Fashionpedia validation images, threshold 0.7: precision 0.837, recall 0.631. Not evaluated on real shop pages yet.

Training

• Base: ustc-community/dfine-nano-coco (Apache-2.0), from D-FINE (Apache-2.0).
• Fine-tuned 6 epochs at 320 px on 13,036 images from Fashionpedia train (shards 0-1).
• Class map from Fashionpedia: upper = {0,1,2,3,4,5,9,12}, lower = {6,7,8}, dress = {10,11}.

Attribution

Fashionpedia annotations are licensed CC BY 4.0 (Jia et al., "Fashionpedia: Ontology, Segmentation, and an Attribute Localization Dataset", ECCV 2020). Fashionpedia images come from sources with their own terms.