Contamination Map Methodology

Vision / Dataset

VisA

10,821 high-resolution color images (9,621 normal, 1,200 anomalous) of 12 objects (PCB1-4, Capsules, Candles, Macaroni1, Macaroni2, Cashew, Chewing gum, Fryum, Pipe fryum) with image- and pixel-level annotations. In the 1-class protocol 90% of normal images form the train set and the remaining 10% of normal images plus all anomalous images form the test set; 2-class high-/low-shot splits are also defined. The released split files reorganise the data into MVTec-AD-style train (normal) / test (normal + anomalous) / ground_truth folders.

Also known as
Visual Anomaly dataset, Visual Anomaly (VisA) Dataset, VisA dataset
Domain
manufacturing inspection
License
CC BY 4.0
Test items from
2022-07-28
Evidence
“arXiv:2207.14315v1 [cs.CV] 28 Jul 2022”
arxiv.org · paper · header · retrieved 2026-10-15
“We release the Visual Anomaly (VisA) Dataset consisting of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains”
arxiv.org · paper · Abstract · retrieved 2026-10-15
Homepage
github.com

Exposure by model

BenchmarkModelLevelReason
Industrial anomaly detection AA-CLIP, trained on MVTec AD (for VisA results) Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection AA-CLIP, trained on VisA Seen, split unclear in fine-tuning data: the dataset was used; the source does not say which part
Industrial anomaly detection AdaCLIP, trained on all 14 datasets (demo model) Test data seen in fine-tuning data: its test split was used for training
Industrial anomaly detection AdaCLIP, trained on MVTec AD + ColonDB Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection AdaCLIP, trained on VisA + ClinicDB Test data seen in fine-tuning data: its test split was used for training
Industrial anomaly detection AdaptCLIP, trained on MVTec AD Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection AdaptCLIP, trained on VisA Test data seen in fine-tuning data: its test split was used for training
Industrial anomaly detection Anomaly-OneVision (Anomaly-OV) Seen, split unclear in fine-tuning data: the dataset was used; the source does not say which part
Industrial anomaly detection AnomalyCLIP, trained on MVTec AD Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection AnomalyCLIP, trained on VisA Test data seen in fine-tuning data: its test split was used for training
Industrial anomaly detection AnomalyDINO Unknown unknown for its base model DINOv2: training data not disclosed in enough detail
Industrial anomaly detection AnomalyGPT, unsupervised on MVTec AD Unknown training data not disclosed in enough detail
Industrial anomaly detection AnomalyGPT, supervised (demo model) Seen, split unclear in fine-tuning data: the dataset was used; the source does not say which part
Industrial anomaly detection AnomalyGPT, unsupervised on VisA Training data seen in fine-tuning data: only the training split was used, not the benchmark's test data
Industrial anomaly detection APRIL-GAN (VAND 2023), trained on MVTec AD Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection APRIL-GAN (VAND 2023), trained on VisA Test data seen in fine-tuning data: its test split was used for training
Industrial anomaly detection OpenAI CLIP Unknown training data not disclosed in enough detail
Industrial anomaly detection DINOv2 Unknown training data not disclosed in enough detail
Industrial anomaly detection DINOv3 (web / LVD-1689M models) Unknown training data not disclosed in enough detail
Industrial anomaly detection FiLo, trained on MVTec AD Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection FiLo, trained on VisA Test data seen in fine-tuning data: its test split was used for training
Industrial anomaly detection WideResNet-50-2 (supervised ImageNet-1k, torchvision) Not seen absent from the complete list of training data
Industrial anomaly detection InCTRL, trained on MVTec AD Not seen absent from the complete list of training data and from its base models' data
Industrial anomaly detection InCTRL, trained on VisA Test data seen in fine-tuning data: the whole dataset was used
Industrial anomaly detection MetaUAS Unknown training data not disclosed in enough detail
Industrial anomaly detection MuSc Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection OpenCLIP (LAION-2B checkpoints) Not seen its test items were published from 2022-07-28, after the model's stated training-data cutoff (2022-01-31)
Industrial anomaly detection OpenCLIP (LAION-400M checkpoints) Not seen its test items were published from 2022-07-28, after the model's stated training-data cutoff (2021-12-31)
Industrial anomaly detection Segment Any Anomaly (SAA / SAA+) Unknown unknown for its base model Segment Anything Model (SAM): training data not disclosed in enough detail
Industrial anomaly detection SAM 2 / SAM 2.1 Unknown training data not disclosed in enough detail
Industrial anomaly detection Segment Anything Model (SAM) Unknown training data not disclosed in enough detail
Industrial anomaly detection SigLIP 2 Unknown training data not disclosed in enough detail
Industrial anomaly detection SigLIP Unknown training data not disclosed in enough detail
Industrial anomaly detection UniVAD Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection VCP-CLIP, trained on MVTec-AD Unknown unknown for its base model OpenAI CLIP: training data not disclosed in enough detail
Industrial anomaly detection VCP-CLIP, trained on VisA Test data seen in fine-tuning data: its test split was used for training
Industrial anomaly detection WinCLIP / WinCLIP+ Not seen absent from the complete list of training data and from its base models' data

Notes

Hosted as a single tar archive on a public AWS S3 URL (no form); the archive file name is VisA_20220922.tar, which may indicate a later re-upload. The 1-class split is defined by split_csv/1cls.csv in the repository.

Sources