Cross-Teaching Segmentation
1 language| Type | Interactive demo |
|---|---|
| Origin | DTU course 02456, Deep Learning |
| Developer | David |
| Written in | PyTorch, segmentation-models-pytorch, timm, TypeScript, Canvas |
| Tests | Predictions, confidence maps, and attention rollouts generated from four shipped checkpoints |
| Built with | DTU 02456 Group 9 project, plus a 2026 Oxford-IIIT Pet redesign |
| Hosted on | David-Internet/demos/crossteach |
| Website | https://david-internet.vercel.app/demos/crossteach |
Cross-Teaching Segmentation is an interactive demonstration of semi-supervised segmentation in which a U-Net and a Vision Transformer grade each other's predictions on unlabeled images.[1] It presents the DTU 02456 Group 9 project from fall 2025 and a later Oxford-IIIT Pet redesign, using four shipped checkpoints rather than simplified stand-ins.
Visitors can pick a pet, compare supervised and cross-taught masks, dim low-confidence pixels, and press teach so that pseudo-labels cross the resolution gap between a 512-pixel U-Net and a 224-pixel ViT.
Overview
The page includes training curves from committed per-epoch metrics, a results table for Oxford Pet, and the original micro-CT story (Dice 0.49 to 0.97 with 22 labeled slices). Encoder activations, attention rollout, and a learning ladder through the course notebooks sit beside the live exchange.
Method
Cross-teaching gates the exchange on confidence: after a two-epoch warmup the confident-image ratio rises and each model trains on the other's pseudo-labels. The demo also documents CrossDetection.py, a detector pair that was written and never run.
Development
Predictions, confidence maps, and attention rollouts were generated by running the public checkpoints over held-out images. Predictions, confidence maps, and attention rollouts generated from four shipped checkpoints.
See also
References
- ^
content/crossteach/README.md, David's Internet.
- Interactive demos
- Machine learning
- 2026 establishments