tae-mo / CRAD

The official repository of Continuous Memory Representation for Anomaly Detection

Continuous Memory Representation for Anomaly Detection (ECCV 2024)

Joo Chan Lee*, Taejune Kim*, Eunbyung Park, Simon S. Woo, Jong Hwan Ko

[Project Page] [Paper(arxiv)]

This repository is the official implementation of Continuous Memory Representation for Anomaly Detection.

Get Started

Environment

Python3.8

Packages:

• torch==1.12.1
• torchvision==0.13.1

Requirements

To install requirements:

pip install -r requirements.txt

Data preparation

1. Download the MVTec AD dataset
2. Construct the data structure as follows:

|-- data
    |-- MVTec-AD
        |-- mvtec_anomaly_detection
            |--bottle
            |--cable
            |-- ...
        |-- train.json
        |-- test.json

Training

To train the model(s) in the paper, run this command:

cd experiments/
bash train.sh 4 0,1,2,3 1111
# bash train.sh <num gpus> <gpu ids> <master port>

Evaluation

To evaluate a trained model, run:

cd experiments/
bash eval.sh 4 0,1,2,3 1111
# bash eval.sh <num gpus> <gpu ids> <master port>

Results

Our model achieves the following performance on MVTec AD:

Platform	GPU	Detection AUROC	Localization AUROC
torch.distributed.launch	4 GPU (NVIDIA RTX A5000 24 GB)	99.3	97.8

BibTeX

@article{lee2024crad,
title={Continuous Memory Representation for Anomaly Detection},
author={Lee, Joo Chan and Kim, Taejune and Park, Eunbyung and Woo, Simon S. and Ko, Jong Hwan},
journal={arXiv preprint arXiv:2402.18293},
year={2024}
}