💻 computer science

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

This paper introduces Continual-MEGA, a large-scale benchmark for generalizable continual anomaly detection that features a diverse dataset including the new ContinualAD, a novel zero-shot generalization scenario, and a unified baseline algorithm demonstrating significant improvements over existing methods in both few-shot detection and pixel-level localization.

Original authors: Geonu Lee, Yujeong Oh, Geonhui Jang, Soyoung Lee, Jeonghyo Song, Sungmin Cha, YoungJoon Yoo

Published 2026-02-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Geonu Lee, Yujeong Oh, Geonhui Jang, Soyoung Lee, Jeonghyo Song, Sungmin Cha, YoungJoon Yoo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the quality control manager for a massive factory that makes everything from toothpaste tubes to toy cars. Your job is to spot defects: a scratch on a car, a dent in a can, or a tear in a fabric.

In the past, you'd hire a specialist for each product. If you started making toothpaste, you hired a "toothpaste expert." If you switched to making cars, you fired the toothpaste expert and hired a "car expert." This is slow, expensive, and wasteful.

This paper introduces a new way to train a single, super-smart AI assistant that can learn to spot defects on any product, one after another, without forgetting how to check the previous ones. They call this "Continual-MEGA."

Here is a breakdown of their work using simple analogies:

1. The Problem: The "Amnesia" Factory

Real-world factories don't stay still. New products arrive every day.

The Old Way: If you retrain your AI every time a new product arrives, it often suffers from "catastrophic forgetting." It's like a student who studies hard for a math test, passes, and then immediately forgets how to do math the moment they start studying history.
The New Challenge: Even worse, what if a defect appears on a product the AI has never seen before? (e.g., a new type of plastic bottle). The AI needs to be smart enough to guess, "This looks wrong," even without ever having seen that specific bottle before. This is called Zero-Shot Learning.

2. The Solution: A New "Gym" for AI (Continual-MEGA)

To test if AI can actually do this, the authors built a massive new training ground called Continual-MEGA.

The Old Gyms: Previous tests were like small, quiet gyms with only a few types of equipment (datasets like MVTec-AD). They were too easy and didn't reflect the chaos of a real factory.
The New Gym (Continual-MEGA): The authors built a giant, chaotic gym. They combined existing datasets with a brand-new dataset they created called ContinualAD.
- ContinualAD is special because it's huge and messy. Instead of taking perfect photos of 10 identical apples, they took photos of apples in different lighting, with different scratches, and even some that were slightly rotten. This forces the AI to learn the concept of a "bad apple" rather than just memorizing a specific picture.

3. The Test: Three Scenarios

They put various AI models through three different "obstacle courses":

The Marathon (Standard Continual Learning): The AI learns to check 85 types of products, then has to learn 60 new types one by one. Can it remember the first 85 while learning the new ones?
The Blind Test (Zero-Shot): After learning the new products, the AI is shown a product it never saw during training (like a specific type of medical device). Can it still spot the defect?
The Stress Test: They removed the "messy" new dataset (ContinualAD) to see if the AI could still generalize. Most AIs failed this, proving that the messy, diverse data was crucial for learning.

4. The Winner: The "Swiss Army Knife" AI

The authors tested many existing AI methods, and most struggled. They performed well on simple tasks but fell apart when the factory got busy and new, weird products arrived.

They proposed their own method, called ADCT (Anomaly Detection across Continual Tasks). Think of this method as a Swiss Army Knife:

The Base: It uses a pre-trained brain (CLIP) that already knows what "normal" and "abnormal" look like in general.
The Attachments (Adapters): Instead of rebuilding the whole brain for every new product, it just snaps on tiny, lightweight attachments (adapters) for each new task.
The Magic Trick (Feature Synthesis): Since they don't have enough photos of every possible defect, the AI imagines what a defect might look like by adding "noise" (random static) to the image features. It's like a chef tasting a soup and saying, "If I add a pinch of salt, it would taste like this," so they don't need to actually cook a thousand different soups to learn the recipe.

5. The Results

The Verdict: Existing methods are like students who memorize answers; they fail when the test changes. The new "Swiss Army Knife" AI is like a student who understands the principles of the subject.
Key Finding: The AI that used the "Swiss Army Knife" approach (lightweight adapters + imagining defects) was the most stable. It didn't forget old products, and it was surprisingly good at spotting defects on products it had never seen before.

Summary

The paper says: "We built a massive, realistic training ground for AI to learn how to spot factory defects on new products without forgetting old ones. We found that current AI is too rigid, but a simple method that combines a pre-trained brain with tiny, flexible add-ons and a bit of 'imagination' works best."

Important Note: The authors strictly tested this on industrial objects (toys, food, electronics, fabrics). They did not claim this works for medical diagnoses or human surveillance, and they emphasized that their data contains no personal information.

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

Problem Statement

Anomaly detection (AD) is critical for quality control in diverse domains, yet existing models struggle in dynamic real-world deployment scenarios. Traditional approaches often assume static, unsupervised, or per-class settings, failing to adapt to sequentially arriving tasks where new object categories or defect types emerge. Retraining models from scratch for every new task is computationally expensive and often infeasible due to data retention policies and privacy regulations. Furthermore, the extreme rarity of anomaly samples makes re-collecting historical defect patterns difficult, leading to catastrophic forgetting.

While recent advancements in Continual Learning (CL) and Zero-Shot Learning (ZSL) offer potential solutions, existing evaluation benchmarks (e.g., MVTec-AD, VisA) are limited in scale and diversity. They do not adequately test a model's ability to generalize to unseen classes after continual adaptation (Continual Zero-Shot Learning, or CZSL) or handle the high variance found in real-world industrial environments. There is a lack of large-scale benchmarks that integrate diverse public datasets with new, high-variance data to rigorously evaluate the generalizability and adaptability of AD systems.

Methodology

The Continual-MEGA Benchmark

The authors introduce Continual-MEGA, a large-scale benchmark designed to evaluate continual and zero-shot capabilities in anomaly detection. It integrates seven existing public datasets (MVTec-AD, VisA, Real-IAD, VIADUCT, BTAD, MPDD) with a newly curated dataset, ContinualAD.

ContinualAD Dataset:

Scale and Diversity: Comprises 30 classes with 14,655 normal and 15,827 anomaly images.
High Variance: Unlike prior datasets where images within a class are visually similar, ContinualAD exhibits significantly higher pixel-level variance. This is achieved by capturing objects under diverse conditions using 10 different devices (ranging from smartphones to tablets) and inducing various defect types (cracks, holes, rot, scratches, etc.).
Annotation: Anomaly regions are annotated at the pixel level using polygon masks.

Evaluation Scenarios:
The benchmark defines three distinct scenarios to test different aspects of model performance:

Scenario 1 (Standard CL): Pre-training on 85 base classes followed by continual learning with 60 new classes introduced in increments (5, 10, or 30 classes per task).
Scenario 2 (CZSL): Pre-training on 58 base classes (excluding MVTec-AD and VisA) followed by continual learning on 60 new classes. Zero-shot generalization is then evaluated on the held-out MVTec-AD and VisA datasets.
Scenario 3 (ContinualAD Effectiveness): Similar to Scenario 2 but excludes the ContinualAD dataset from the training stream to specifically measure the impact of the new dataset on zero-shot generalization.

Metrics:
Performance is measured using Average Accuracy (ACC) based on Image-level AUROC and Pixel-level Average Precision (AP), alongside a Forgetting Measure (FM) to quantify performance degradation on previous tasks.

Proposed Baseline: ADCT

To address the benchmark, the authors propose Anomaly Detection across Continual Tasks (ADCT), a baseline method built upon the CLIP (Contrastive Language-Image Pre-training) framework. Key components include:

Mixture-of-Experts (MoE) Adapters: Lightweight MLP adapters are inserted into four specific blocks of the CLIP visual encoder. Separate adapters are trained for each task (and the base set). During inference, an average of all adapters ( $A_{avg}$ ) is used to preserve knowledge across tasks.
Synthetic Feature Generation: To overcome data scarcity and improve pixel-level localization, the method generates synthetic anomaly features by adding Gaussian noise to the visual features before passing them through the adapters. This allows the model to learn diverse anomaly patterns without requiring extensive real anomaly data.
Text Prompts: Generalized text prompts are used to extract text features for both normal and anomalous states, facilitating zero-shot capabilities.

Key Contributions

Continual-MEGA Benchmark: The introduction of a novel, large-scale benchmark that combines existing public datasets with the high-variance ContinualAD dataset, significantly expanding the volume and diversity of data available for continual AD research.
Comprehensive Evaluation: Extensive evaluation of various state-of-the-art AD methods (including VLM-based, continual learning-specific, and only-normal sample approaches) reveals substantial room for improvement, particularly in pixel-level defect localization and generalizability to unseen classes.
ADCT Baseline: The proposal of a simple yet effective baseline method that combines lightweight MoE-style adapters with feature synthesis. This approach demonstrates that minimal adaptation to pre-trained CLIP, rather than extensive fine-tuning, can yield robust performance in both continual and zero-shot settings.

Results

Extensive experiments on the Continual-MEGA benchmark yield three primary findings:

Limitations of Existing Methods: Current AD methods exhibit significant performance drops, especially in pixel-level localization, when evaluated under the proposed large-scale continual settings. Methods that perform well on standard benchmarks (like MVTec-AD) often fail to generalize to the diverse conditions in Continual-MEGA.
Effectiveness of ContinualAD: The inclusion of the ContinualAD dataset in the training stream (Scenario 2 vs. Scenario 3) significantly improves zero-shot generalization performance for most methods. Conversely, excluding it leads to a noticeable degradation in generalization, highlighting its value for training robust models.
Superiority of ADCT: The proposed ADCT baseline achieves overall superior performance compared to other methods across all scenarios. It maintains high accuracy while minimizing forgetting (low FM). The ablation study confirms that the combination of adapters, synthetic feature generation, and the mixture-of-experts inference scheme is crucial for balancing stability and plasticity. Specifically, synthetic features are shown to be vital for pixel-level performance, while the MoE adapter strategy is key to reducing catastrophic forgetting.

Significance and Claims

The paper posits that Continual-MEGA establishes a new agenda for the anomaly detection field by addressing the gap between static benchmark evaluations and the dynamic realities of industrial deployment. The authors claim that their work highlights the practical necessity of models that can incrementally adapt to new tasks while generalizing to entirely unseen classes (CZSL).

The significance of the proposed baseline (ADCT) lies in its demonstration that lightweight adaptation combined with feature synthesis is a promising direction for designing generalizable continual frameworks. The results suggest that excessive adaptation or heavy guidance to pre-trained models like CLIP may lead to overfitting on previously seen objects, whereas a strategy that preserves the pre-trained knowledge while synthesizing diverse anomaly features offers a more stable and generalizable solution.

The authors conclude that their work contributes to manufacturing sustainability and reproducibility by reducing the computational cost and carbon footprint associated with retraining deep learning models in dynamic environments, without introducing ethical risks related to personal data or surveillance.