Abstract
Anomaly detection (AD) is often focused on detecting anomaly areas forindustrial quality inspection and medical lesion examination. However, due tothe specific scenario targets, the data scale for AD is relatively small, andevaluation metrics are still deficient compared to classic vision tasks, suchas object detection and semantic segmentation. To fill these gaps, this workfirst constructs a large-scale and general-purpose COCO-AD dataset by extendingCOCO to the AD field. This enables fair evaluation and sustainable developmentfor different methods on this challenging benchmark. Moreover, current metricssuch as AU-ROC have nearly reached saturation on simple datasets, whichprevents a comprehensive evaluation of different methods. Inspired by themetrics in the segmentation field, we further propose several more practicalthreshold-dependent AD-specific metrics, ie, m$F_1$$^{.2}_{.8}$,mAcc$^{.2}_{.8}$, mIoU$^{.2}_{.8}$, and mIoU-max. Motivated by GAN inversion'shigh-quality reconstruction capability, we propose a simple but more powerfulInvAD framework to achieve high-quality feature reconstruction. Our methodimproves the effectiveness of reconstruction-based methods on popular MVTec AD,VisA, and our newly proposed COCO-AD datasets under a multi-class unsupervisedsetting, where only a single detection model is trained to detect anomaliesfrom different classes. Extensive ablation experiments have demonstrated theeffectiveness of each component of our InvAD. Full codes and models areavailable at https://github.com/zhangzjn/ader.