AnomalyAgent studies how to combine pretrained visual-anomaly detectors when their predictions disagree. Rather than apply one fixed averaging rule to every image, the framework treats fusion as a constrained choice among several actions. The target problem is anomaly detection where abnormal training examples can be scarce.
Align scores before combining them
The AnomalyAgent paper first aligns heterogeneous image scores and anomaly maps. It then summarizes disagreement at both the image and region levels.
Those levels answer different questions. An image-level score indicates whether an object may be abnormal, while a dense map identifies where the detector sees a problem. Detectors trained with different objectives can disagree about either. A useful ensemble needs to distinguish those conflicts rather than hide them behind one aggregate confidence.
Limit the language model’s choices
The framework defines score actions such as consensus, robust aggregation and detector weighting. It separately defines map actions, including geometric aggregation and disagreement correction. A constrained vision-language model inspects high-conflict samples and selects a valid score action and map action.
This is a selection role over defined operations, not an invitation for the model to invent an unrestricted anomaly map. The distinction makes the fusion procedure more concrete to evaluate, although it does not remove the possibility of choosing an unhelpful action.
Read the benchmark claim precisely
The authors evaluate MVTec AD and VisA as primary benchmarks, alongside seven additional public benchmarks and a large-scale score evaluation on Real-IAD. They report consistent improvement over equal-weight fusion and the best mean image AUROC and pixel average precision among compared fusion methods on the two primary benchmarks.
That comparison concerns the tested fusion methods and datasets, not every possible anomaly-detection system. For an applied evaluation, useful checks include whether disagreement selection adds cost and whether improvements persist on the actual defect types and image conditions. The paper’s contribution is an adaptive way to combine existing detectors, with the results attributed to its reported experiments.
Comments