zhaolutuan/AnomalyCoT

Anomaly detection tasks in real-world settings are typically complex and variable, and various types of defects can occur in different products. Therefore, the datasets we construct are required to cover multiple scenarios and defect types of anomaly detection. We initially acquired and sampled from nine publicly available IAD datasets. To further expand the scenarios, we also manually gathered data from four scenarios applicable to IAD tasks to form several independent datasets and eliminated low-quality data through manual examination. MVTec AD is one of the most renowned IAD datasets, containing high-resolution images of various objects and corresponding to different types of defects. The VisA \cite{35} dataset further enhances the complexity of scenarios and the data scale, including different instances of the same type of objects. GoodsAD \cite{36} offers the possibility for intelligent applications of anomaly detection by establishing the first IAD commodity dataset. MVTec LOCO AD \cite{37} is dedicated to exploring logical anomalies and constructs a novel dataset covering both structural and logical anomalies. We also adopted Defect Spectrum \cite{38}, which effectively optimizes based on multiple key IAD datasets.

In industrial fields with less attention, MPDD \cite{39} and NanoTwice \cite{40} are respectively designed for anomaly detection tasks in the production processes of painted metal parts and nanofiber materials. MSD \cite{41} is a potential dataset for surface defect detection of smartphone screens. MIAD is the largest component of the dataset in this paper, containing over 100K high-resolution outdoor IAD images. The large-scale, high-quality data make a significant contribution to the application of IAD tasks in practical environments. Additionally, we collected IAD data related to wood, fabric, car, and road cracks, screened them for high-quality samples, and then manually classified the unlabeled data to build four independent datasets. These independent datasets fill the voids in certain scenarios of the current task. We integrated the aforementioned processed datasets, precisely located all anomalous regions, and classified the defects within them through manual annotation, then verified the correctness of the labels through manual inspection, and finally manually corrected the incorrect labels. A total of 10 individuals participated in this labeling process, which accumulated to 230 hours.