Boosting Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Adversarial Learning

PyTorch implementation of Boosting Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Adversarial Learning

arxiv

Dataset

Please download these four datasets Chest X-rays OCT ISIC2018 Br35H MVTec and place them in ./data/ as follows:

|-- ChestXRay2017
|   |-- train
|   |   |-- NORMAL
|   |   |-- PNEUMONIA
|   |
|   |-- test
|
|-- OCT2017
|   |-- train
|   |   |-- CNV
|   |   |-- DME
|   |   |-- DRUSEN
|   |   |-- NORMAL
|   |
|   |-- test
|
|-- ISIC2018
|   |-- ISIC2018_Task3_Training_Input
|   |-- ISIC2018_Task3_Training_GroundTruth
|   |-- ISIC2018_Task3_Test_Input
|   |-- ISIC2018_Task3_Test_GroundTruth
|   |-- ISIC2018_Task3_Validation_Input
|   |-- ISIC2018_Task3_Validation_GroundTruth
|
|-- Br35H
|   |-- yes
|   |-- no
|
|-- mvtec
|   |-- carpet
|   |-- grid
|   |-- ...
|
|--visa
|   |--candle
|   |--capsules
|   |--...
|

Installation

Please install the dependency packages using the following command by pip:

pip install -r requirements.txt

Taining

You can run the corresponding script in the dir ./script/ to run the certain dataset after setting the appropriate parameters. For normal data, in Chest X-rays, OCT, Br35H and ISIC2018, the normal category is 0, in MVTec, the normal class can be carpet, grid, and other 13 classes.

To run on the Chest X-rays dataset, you can run the script:

bash ./script/xray.sh