A curated list of dataset for industrial anomaly detection.
- VisA : The VisA dataset contains 12 subsets corresponding to 12 different objects. There are 10,821 images with 9,621 normal and 1,200 anomalous samples. Four subsets are different types of printed circuit boards (PCB) with relatively complex structures containing transistors, capacitors, chips, etc. For the case of multiple instances in a view, there are four subsets: Capsules, Candles, Macaroni1, and Macaroni2. Instances in Capsules and Macaroni2 largely differ in locations and poses. Moreover, there are four subsets, including Cashew, Chewing gum, Fryum, and Pipe fryum, where objects are roughly aligned. The anomalous images contain various flaws, including surface defects such as scratches, dents, color spots, or cracks, and structural defects like misplacement or missing parts. [link]
- MVTec-AD: MVTec-AD contains 5,354 high-resolution images divided into 10 different objects and 5 texture categories. MVTec-AD consists of 3629 images for training and validation and 1,725 images for testing. The training set contains only normal images, while the test dataset contains both normal and anomalous images. [link]
- MVTec-LOCO AD: The MVTec LOCO AD dataset is intended for the evaluation of unsupervised anomaly localization algorithms. The dataset includes both structural and logical anomalies. It contains 3,644 images from five different categories. The dataset also includes pixel-precise ground truth data for each anomalous region. [link]
- **Real-IAD **: Real-IAD is a large-scale, real-world, and multi-view industrial anomaly detection dataset, which contains 151,050 high-resolution images of 30 different objects (99,721 normal images and 51,329 anomaly images). It has a larger range of defect area and ratio proportions. [link]
- BTAD: The BTAD dataset contains a total of 2,540 real-world images of 3 industrial products showcasing body and surface defects, in which Products 1, 2, and 3 have 400, 1,000, and 399 train images, respectively. All classes in this dataset belong to textures. [link]
- DGAD: DAGM consists of ten texture classes with 15,000 normal images and 2,100 abnormal images. Various defects that are visually close to the background, such as scratches and specks, constitute anomalous samples. We still perform the unsupervised paradigm, where the training set contains only normal samples. [link]
- MPDD: MPDD contains 6 classes of metal parts, focusing on defect detection during the fabrication of painted metal parts. Its training set is composed of 888 normal samples without defects, and the test set is composed of 458 samples either normal or anomalous. In particular, samples in MPDD have non-homogeneous backgrounds with diverse spatial orientations, different positions, and various light intensities, leading to greater challenges in anomaly detection. [link]
- MTD: MTD includes 1344 images, and ROIs of the surface of magnetic tiles are cropped. Image files and pixel-level labels are separated into six datasets according to different defect types: Blowhole, Crack, Fray, Break, Uneven (grinding uneven), and Free (no defects). [Link]
- PKU-Market-Phone: This dataset contains 3 types of surface defects: Oil, Scratch, and Stain. It consists of 1,200 images and 400 images for each defect. The defects are made by ourselves. The images are collected by an industrial camera, and the resolution is 1920×1080. The dataset is randomly divided into train:val:test=6:2:2. The dataset format is PASCAL VOC. [Link]
- PKU-GoodsAD: The GoodsAD dataset comprises 6 categories with 3136 images for training and 2988 images for testing. The dataset contains a total of 484 goods. Each category contains several common defects such as surface damage, deformation and opened. All images are acquired with 3000 ×3000 high-resolution. [Link]
- InsPLAD: InsPLAD1 is a power line asset inspection in-the-wild dataset that offers multiple computer vision challenges, one being anomaly detection in power line components called InsPLAD-fault. Its data are real-world unmanned aerial vehicle images of operating power line transmission towers. It contains five power line object categories with one or two types of anomalies for each class, resulting in 11,662 images, of which 402 are samples of defective objects annotated on the image level. [Link]
- VISION: The VISION Datasets is a collection of 14 industrial inspection datasets, designed to explore the unique challenges of vision-based industrial inspection. These datasets cover a wide range of manufacturing processes, materials, and industries. VISION V1 dataset includes a total of 18,422 images, of which 4,165 images are annotated. The VISION V2 dataset with full annotation is released with 13804 annotated images. [Link]
- MAD:The MAD dataset containing 4,000+ highresolution multi-pose views RGB images with camera/pose information of 20 shape-complexed LEGO animal toys for training, as well as 7,000+ simulation and real-world collected RGB images (without camera/pose information) with pixel-precise ground truth annotations for three types of anomalies in test sets. Note that MAD has been further divided into MAD-Sim and MAD-Real for simulation-to-reality studies to bridge the gap between academic research and the demands of industrial manufacturing.[Link]
- Defect Spectrum: The Defect Spectrum dataset offers precise, semantics-abundant, and large-scale annotations for a wide range of industrial defects. This dataset is an enhancement over existing benchmarks(including MVTec-AD, VISION, DAGM, and Cotton-Fabric.), providing refined annotations and introducing detailed semantic layers, allowing for the distinction between multiple defect types within a single image. [Link]
- MVTec 3D-AD: MVTec 3D-AD is a 3D anomaly detection dataset comprising over 4,000 RGB images and the corresponding high-resolution 3D point cloud data. Each of the 10 sub-categories of MVTec-3D AD is divided into a defect-free training set (2,656 training samples) and a test set (1,137 testing samples) containing various kinds of defects. [Link]
- Anomaly-ShapeNet: Anomaly-ShapeNet comprises a total of 1,600 samples which are distributed across 40 distinct categories. There are six kinds of anomalies, including bulge, concavity, crack, holes, and broken. Each training set for a category contains only four samples. Each test set for a category contains normal and various defective samples. [Link]
- Real3D-AD: Real3D-AD comprises a total of 1,254 samples that are distributed across 12 distinct categories. These categories include Airplane, Car, Candybar, Chicken, Diamond, Duck, Fish, Gemstone, Seahorse, Shell, Starfish, and Toffees. Each training sample is an absence of blind spots, and a realistic, high-accuracy prototype. There are 48 samples in the training set and 1,206 samples for testing. [Link]
More details can be found in [Link]