README.md

metadata

license: cc-by-nc-4.0
task_categories:
  - image-segmentation
tags:
  - anomaly-detection
  - continual-learning
  - benchmark

Continual-MEGA: A Large-scale Benchmark for Generalizable Continual Anomaly Detection

This repository contains the dataset for Continual-MEGA, a new benchmark for continual learning in anomaly detection, introduced in the paper Continual-MEGA: A Large-scale Benchmark for Generalizable Continual Anomaly Detection.

Continual-MEGA aims to better reflect real-world deployment scenarios. It features a large and diverse dataset that significantly expands existing evaluation settings by combining carefully curated existing datasets with the newly proposed ContinualAD dataset. The benchmark also proposes a novel scenario for measuring zero-shot generalization to unseen classes, particularly focusing on pixel-level defect localization.

For the associated evaluation code, checkpoint files, and further details, please refer to the GitHub repository: https://github.com/Continual-Mega/Continual-Mega-Neurips2025

Dataset Structure

The Continual-MEGA benchmark dataset combines data from various sources, structured as follows:

data/
├── continual_ad/              # Our proposed ContinualAD dataset
├── mvtec_anomaly_detection/   # MVTec-AD dataset
├── VisA_20220922/             # VisA dataset
├── VIADUCT/                   # VIADUCT dataset 
├── Real-IAD-512/              # RealIAD dataset (512 size)
├── MPDD/                      # MPDD dataset
└── BTAD/                      # BTAD

Sample Usage (Evaluation)

The evaluation code for the Continual-MEGA benchmark is available in the associated GitHub repository. After cloning the repository and setting up, you can run the following commands:

Continual Settings Evaluation

sh eval_continual.sh

Zero-Shot Generalization Evaluation

sh eval_zero.sh

For detailed setup instructions, including downloading CLIP pretrained weights and specific checkpoint files, please visit the official GitHub repository.