Published March 22, 2024 | Version v1
Dataset Open

Long-Tailed Anomaly Detection (LTAD) Dataset

  • 1. University of California San Diego
  • 2. Mitsubishi Electric Research Laboratories
  • 3. ROR icon University of California, San Diego



Anomaly detection (AD) aims to identify defective images and localize their defects (if any). Ideally, AD models should be able to: detect defects over many image classes; not rely on hard-coded class names that can be uninformative or inconsistent across datasets; learn without anomaly supervision; and be robust to the long-tailed distributions of real-world applications. To address these challenges, we formulate the problem of long-tailed AD by introducing several datasets with different levels of class imbalance for performance evaluation.

To encourage more follow up works on long-tailed AD, we are publicly releasing the dataset split used in our paper (“Long-Tailed Anomaly Detection with Learnable Class Names” by Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos, CVPR 2024).

Files in the unzipped folder:

1.      ./ This Markdown file

2.      ./dataset_split: Folder contains long-tail splits from three datasets. See below for details.


At a Glance

  • The size of the unzipped dataset is ~16MB
  • Three datasets are used in this project, including [MVTec](, [VisA]( and [DAGM]( Please download the datasets from their original repositories.
  • The dataset split provided in this folder is organized as follows:
    |-----|-- exp
    |-----|-----|----- 100
    |-----|-----|----- |-----test.json
    |-----|-----|----- |-----train.json
    |-----|-----|----- 200
    |-----|-- step
    |-----|-- ...
  • Each long-tailed dataset split contains a subfolder ``imbalance_type/imbalance_factor", where imbalance type can be [exponential (exp), step, reverse exponential (exp_reverse), reverse step (step_reverse)]. The definition of imbalance type and imbalance factor can be found in our paper. Each subfolder contains two json files, one for training and the other for testing.
  • Each entry in the json file contains the meta information of an image and is similar to
    {"filename": "candle/test/bad/000.JPG", "label": 1, "label_name": "defective", "clsname": "candle", "maskname": "candle/ground_truth/bad/000.png"}
    - filename: location of the input image in the dataset
    - label: indicates whether the input image is normal (labeled as 0) or defective (labeled as 1)
    - label name: can be "good" or "defective"
    - clsname: class name of the input image
    - maskname (optional): location of the binary image that indicates the defect region. This is only available for test.json, because there is no defect image during training.


If you use the LTAD dataset in your research, please cite our contribution:

author = {Ho, Chih-Hui and Peng, Kuan-Chuan and Vasconcelos, Nuno},
title = {Long-Tailed Anomaly Detection with Learnable Class Names},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2024}


The LTAD dataset is released under CC-BY-SA-4.0 license. For the images in the MVTec, VisA, and DAGM datasets, please refer to their websites for their copyright and license terms.

Created by Mitsubishi Electric Research Laboratories (MERL), 2023-2024
SPDX-License-Identifier: CC-BY-SA-4.0


Files (441.1 kB)

Name Size Download all
441.1 kB Preview Download