Reiji Saito

Real-world anomaly detection isn't black and white: this paper introduces new scenarios and a method to handle the gray areas where "normal" is ambiguous.

Mixing labeled data directly into the ClassMix semi-supervised segmentation pipeline yields surprisingly strong gains, suggesting that even imperfectly aligned labeled data is more valuable than noisy pseudo-labels.

Domain generalization gets a boost: disentangling category information from domain differences and bridging the remaining gap with a SQ-VAE leads to improved segmentation accuracy on unseen domains.