Patch SVDD for Image anomaly detection. Paper: https://arxiv.org/abs/2006.16067 (published in ACCV 2020).

An input image and a generated anomaly map for 'wood' class.

The code runs on python 3.7, pytorch 1.3.1, and torchvision 0.4.2.

• Install required libraries.

• Download MVTec AD dataset: Download
• Untar the 'mvtec_anomaly_detection.tar.xz' file.

Set the DATASET_PATH to the root path of the downloaded MVTec AD dataset.

python main_train.py --obj=bottle --lr=1e-4 --lambda_value=1e-3 --D=64

• obj denotes the name of the class out of 15 MVTec AD classes.
• lr denotes the learning rate of Adam optimizer.
• lambda_value denotes the value of 'lambda' in Eq. 6 of the paper.
• D denotes the number of embedding dimension (default to 64).

python main_evaluate.py --obj=bottle

• obj denotes the name of the class.

The script loads the trained encoder saved in ckpts/ directory. Note that the same evaluation procedure is performed at every training epoch in Step 2.

For a quick evaluation, trained encoders for cable and wood classes are included. Training (Step 2) can be skipped for those classes.

python main_visualize.py --obj=bottle

• obj denotes the name of the class.

The script generates and saves anomaly maps for all the test images in the obj class. The genereated maps are saved in anomaly_maps/obj directory.