This is the source code of our work "Cognitive Memory-Augmented Network for Visual Anomaly Detection".

We recommended the following dependencies:

• python 3.6
• numpy 1.16.2
• torch 1.3.1
• ptable 0.9.2
• scikit-learn 0.22.2
• tqdm 4.42.1
• scipy 1.1.0

• MNIST and CIFAR-10 will be downloaded for you by torchvision.

• Checkpoints for mnist and cifar10 datasets are available here (code: z9jt).

• The folder structure of checkpoints is listed as follow.

  |-- checkpoints',
    |-- mem_cifar',
    |   |-- 0.pkl',
    |   |-- 1.pkl',
    |   |-- 2.pkl',
    |   |-- 3.pkl',
    |   |-- 4.pkl',
    |   |-- 5.pkl',
    |   |-- 6.pkl',
    |   |-- 7.pkl',
    |   |-- 8.pkl',
    |   |-- 9.pkl',
    |-- mem_mnist',
        |-- 0.pkl',
        |-- 1.pkl',
        |-- 2.pkl',
        |-- 3.pkl',
        |-- 4.pkl',
        |-- 5.pkl',
        |-- 6.pkl',
        |-- 7.pkl',
        |-- 8.pkl',
        |-- 9.pkl',

Once your setup is complete, running tests is as simple as running test.py. Usage:

usage: python test.py [-h] [--dataset DATASET] [--path PATH]
                      [--checkpoints CHECKPOINTS]

optional arguments:
  -h, --help  show this help message and exit
  --dataset, DATASET  The name of the dataset to perform tests. 
                      Choose among `mnist`, 'cifar' (default: None)
  --path, PATH        The file path of the dataset to perform tests.
                      (default: None)
  --checkpoints, CHECKPOINTS
                      The checkpoints path of the dataset to perform tests.
                      (default: None)

Example:

Testing on MNIST dataset.

python test.py --dataset 'mnist' --path 'dataset_path' --checkpoints 'checkpoints/mem_mnist'

Testing on CIFAR10 dataset.

python test.py --dataset 'cifar10' --path 'dataset_path' --checkpoints 'checkpoints/mem_cifar10'

Experimental details is explained in our paper "Cognitive Memory-Augmented Network for Visual Anomaly Detection". We list some of the experimental results as follows.

We have performed the proposed method on video dataset. The AUC results of different methods are listed as follow:

And the examples of score results, obtained by the proposed CMAN method on UCSD Ped2 and ShanghaiTech datasets, are listed as follow:

We have also performed the proposed method on image dataset. The AUC results of different methods are listed as follows:

We have also conducted a visual inspection experiment on the MVTec dataset, compared with stateof-the-art methods. The performance of the proposed method on visual inspection task is listed as follow:

This code follows the basic structure from Latent Space Autoregression for Novelty Detection and memae-anomaly-detection.