This is the official repository to the WACV 2021 paper "Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows" by Marco Rudolph, Bastian Wandt and Bodo Rosenhahn.

If the only reason you ended up here is because you made a typo on 'different' - what was our intention - here is a shortened summary: We introduce a method that is able to find anomalies like defects on image data without having some of them in the training set. Click here to watch our short presentation from WACV.

You will need Python 3.6 and the packages specified in requirements.txt. We recommend setting up a virtual environment with pip and installing the packages there.

Install packages with:

$ pip install -r requirements.txt

All configurations concerning data, model, training, visualization etc. can be made in config.py. The default configuration will run a training with paper-given parameters on the provided dummy dataset. This dataset contains images of 4 squares as normal examples and 4 circles as anomaly.

To start the training, just run main.py! If training on the dummy data does not lead to an AUROC of 1.0, something seems to be wrong. Please report us if you have issues when using the code.

The given dummy dataset shows how the implementation expects the construction of a dataset. Coincidentally, the MVTec AD dataset is constructed in this way.

Set the variables dataset_path and class_name in config.py to run experiments on a dataset of your choice. The expected structure of the data is as follows:

train data:

        dataset_path/class_name/train/good/any_filename.png
        dataset_path/class_name/train/good/another_filename.tif
        dataset_path/class_name/train/good/xyz.png
        [...]

test data:

    'normal data' = non-anomalies

        dataset_path/class_name/test/good/name_the_file_as_you_like_as_long_as_there_is_an_image_extension.webp
        dataset_path/class_name/test/good/did_you_know_the_image_extension_webp?.png
        dataset_path/class_name/test/good/did_you_know_that_filenames_may_contain_question_marks????.png
        dataset_path/class_name/test/good/dont_know_how_it_is_with_windows.png
        dataset_path/class_name/test/good/just_dont_use_windows_for_this.png
        [...]

    anomalies - assume there are anomaly classes 'crack' and 'curved'

        dataset_path/class_name/test/crack/dat_crack_damn.png
        dataset_path/class_name/test/crack/let_it_crack.png
        dataset_path/class_name/test/crack/writing_docs_is_fun.png
        [...]

        dataset_path/class_name/test/curved/wont_make_a_difference_if_you_put_all_anomalies_in_one_class.png
        dataset_path/class_name/test/curved/but_this_code_is_practicable_for_the_mvtec_dataset.png
        [...]

Some code of the FrEIA framework was used for the implementation of Normalizing Flows. Follow their tutorial if you need more documentation about it.

Please cite our paper in your publications if it helps your research. Even if it does not, you are welcome to cite us.

@inproceedings { RudWan2021,
author = {Marco Rudolph and Bastian Wandt and Bodo Rosenhahn},
title = {Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows},
booktitle = {Winter Conference on Applications of Computer Vision (WACV)},
year = {2021},
month = jan
}

Another paper link because you missed the first one:

• Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows

This project is licensed under the MIT License.