Prompt Description

APPROVED. The diagram illustrates a dual-branch semi-supervised learning framework designed for realistic long-tailed data in computer vision. The framework comprises two data streams: Labeled Data (Long-Tailed) and Unlabeled Data (Unknown Distribution). Both streams are processed by a shared feature extractor (Backbone Network). Post-feature extraction, the features diverge into two parallel branches: a Biased Predictor, which generates pseudo-labels for unlabeled data and is adaptive to the unlabeled data distribution, and a Balanced Predictor, trained to encourage balanced decision boundaries and used for final inference. A Learning-Status-Guided Feature Diffusion Suppression module is positioned between the backbone and both predictors, connected to the feature space. This module suppresses feature expansion for well-learned classes while preserving exploration for under-learned classes.