MvTec-Anomaly-Detection

Dataset Download

Download the MVTec AD dataset from the following link:

MVTec AD Dataset

The official website of the dataset is:

MVTec Official AD Dataset

Learning of Convolutional Autoencoder

Learning of Convolutional Autoencoder for Anomaly Detection in MVTec AD Dataset up to 100 epochs.

Usage

Training

Input arguments for training are as follows:

python train.py 
--data-path <path_to_dataset> 
--batch-size <batch_size>
--num-workers <number_of_workers>
--lr <learning_rate>
--epochs <number_of_epochs>
--device <device>
--n-save <num_of_images_per_class_to_save>
--emb-dim <embedding_dimension>

Default values used for training on an NVIDIA RTX 3080 + i9-14900k are:

python train.py
--data-path <path_to_dataset>
--batch-size 16
--num-workers 4
--lr 1e-3
--epochs 100
--device cuda
--n-save 5
--emb-dim 1024

Future Work

• More experiments with Convolutional Autoencoder
• Evaluation on detection of anomalies

References

• The MVTec Anomaly Detection Dataset [Bergmann et al., in: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9584-9592, 2019]