metadata
title: Selective Segmentation (MC Dropout)
emoji: 🐾
colorFrom: indigo
colorTo: yellow
sdk: static
app_file: index.html
pinned: false
license: mit
short_description: Pet segmentation that abstains when unsure (in-browser)
Selective segmentation with MC Dropout
U-Net (7.8M parameters) trained on Oxford-IIIT Pet to segment pet / background / border. MC Dropout (20 passes) measures per-pixel uncertainty and the model abstains on the most uncertain pixels: on the 3,669-image test set, pixel error drops from 8.13% to 2.18% at 80% coverage and selective mIoU rises from 0.777 to 0.882. Thresholds and the temperature were chosen on validation data only. Runs entirely in the browser with ONNX Runtime Web; dropout masks are drawn in JavaScript and fed to the exported graph.