Conceptual
Login

Aleatoric Uncertainty and Energy-Based Losses for Semi-Supervised Semantic Segmentation

A training scheme (DUEB) for semi-supervised semantic segmentation that adds two losses to a dual-branch union-intersection pseudo-label framework: a heteroscedastic aleatoric loss, where each branch predicts per-pixel variance that distorts the logits to model inherent data noise at boundaries and low-contrast regions, and an energy-based loss that reinterprets the discriminative classifier as an energy-based model of the joint input-label distribution, adding generative regularization. Both losses are applied with pseudo-labels on unlabeled images and ground truth on labeled ones.