Inputs from 01–02
03 · Learning and regularization
Use the updated model
01 · Source imagesRendering label: bicycle
02 · Target pseudo-labelStreet image: motorcycle
Source · joint learningPseudo-label generation ↔ optimizationThe source-domain training couples pseudo-label generation and optimization to improve transfer across domains.
Target · regularized learningInitial training · regularized lossThe introduction describes replacing cross-entropy loss with a regularized loss during initial training, minimized on the target images.
Updated classifierReady to predict again
02 · Turn predictions into labelsLater pseudo-labels
04 · Classify the new sceneImage → category prediction
Continue learning
After adaptation