Project Objective

MoViAD is an Open Source library for easy and modular Visual Anomaly Detection, built for industrial and research purposes. The library contains some State-of-the-Art models with their trainers, standard datasets, evaluator for calculating standard metrics, and a feature extractor. The library structure is totally modular, allowing an easy isolation of the components needed for your project. The library will support different scenarios:

• Standard anomaly detection (model training and evaluation on a given category) ✅
• Contaminated anomaly detection (model training ed evaluation on a given category with a contaminate training set) ✅
• Continual anomaly detection (model training and evaluation considering a stream of tasks) 🚧 (Work in progress)
• Few shot anomaly detection 🚧 (Work in progress)

How to Install

Inside the main repository directory, run the following command:

Editable mode (if you need to work on the code):

pip install -e ./

Fixed mode (if you just want to use the code):

pip install ./

How to use the library

The library follows this structure:

• inside the /models/model_name directory is present the code of anomaly detection models
• inside the /trainers directory is present the code for training the anomaly detection models
• inside the /datasets directory is present the code for the anomaly detection datasets that must be used
• inside the /utilies directory is present the code for anomaly detection utilities

Execution example

Inside the /main_scripts directory are present some execution scripts for training and testing the AD models.

For example, for training patchcore:

python main_scripts/main_patchcore.py --mode train --dataset_path /home/datasets/mvtec --category pill --backbone mobilenet_v2 --ad_layers features.4 features.7 features.10 --device cuda:0 --save_path ./patch.pt 

For every main script all its parameters are documented.

AD Models

• PatchCore: paper, code
• CFA paper, code
• Student-Teacher Feature Pyramid: paper, code
• PaSTe: paper
• PaDiM: paper with code
• FastFlow: paper
• GANomaly: paper
• SuperSimpleNet: paper
• RD4AD: paper

Feature Extraction Backbones

• MobileNet: link
• PhiNet: link
• MicroNet: link
• MCUNet: link
• WideResnet

Datasets

• MVTecAD: link
• VisA: link
• RealIAD: paper
• MIIC Dataset: link

Contribute

If you want to contribute to the repository, follow the present code structure:

• inside the /models/model_name directory put the code for possible new anomaly detection models
• inside the /trainers directory put the code for training an anomaly detection model
• inside the /datasets directory put the code for possible new anomaly detection datasets that must be used

Every contribution must be open with a pull request.

Citations

If you use MoViAD in your work, please cite us! 🤗

Works that uses MoViAD:

• "PaSTe: Improving the efficiency of visual anomaly detection at the edge", Manuel Barusco, Francesco Borsatti, Davide Dalle Pezze, Francesco Paissan, Elisabetta Farella, Gian Antonio Susto. paper
• "From Vision to Sound: Advancing Audio Anomaly Detection with Vision-Based Algorithm", Manuel Barusco, Francesco Borsatti, Davide Dalle Pezze, Francesco Paissan, Elisabetta Farella, Gian Antonio Susto. paper