17 Jul 2025
Researchers at UNSW Sydney developed Class-aware Contrastive Learning (CCL), a modular framework that mitigates inter-class confusion in multi-class anomaly detection by integrating local and global contrastive losses with existing reconstruction-based models. This approach achieved an Image-level AUROC of 90.6% across 60 object categories and demonstrated comparable performance using pseudo-class labels, making it suitable for truly unsupervised scenarios.
06 Dec 2024
The MANTA dataset, developed by researchers from UNSW Sydney and collaborators, introduces the first large-scale, multi-view, visual-text resource for anomaly detection in tiny objects (4-20 mm³), featuring over 137,000 multi-view images across 38 categories. This dataset addresses challenges like object heterogeneity and unpredictable poses, demonstrating that multi-view approaches significantly improve anomaly detection, with an average I-AUROC of 91% on its visual tasks.
There are no more papers matching your filters at the moment.
