In this paper, a novel feature reconstruction and disruption model (FRD-UVAD) is proposed for effective feature refinement and better pseudo-label generation in fully UVAD, based on cascade cross-attention transformers, a latent anomaly memory bank and an auxiliary scorer.
