Old is Gold: Redefining the Adversarially Learned One-Class Classifier Training Paradigm (CVPR 2020)
CVPR Presentation || Paper || CVPR CVF Archive || Supp material zip file || Ped2 Results Video
[17.6.2020] For the time being, test code (and some trained models) are being made available. Training code will be uploaded in some time.
- Python3.5
- Pytorch0.4.0 (see: pytorch installation instuctions)
- Download trained generator and discriminator models from here and place inside the directory ./models/
- Download datasets here and place test images in the subdirectories of ./data/test/
- Example:
- All images from inlier class (*.png) should be placed as ./data/test/0/*.png
- Similarly, all images from outlier class (* *.png) should be placed as ./data/test/1/* *.png
- Example:
- run test.py
The models provided are trained on the training set of '0' class in MNIST dataset. For evaluation, the test dataset provided contains all test images from class '0' as inliers, whereas 100 images each from all other classes as outliers.
If you have any query, please feel free to contact Zaigham through mzz.pieas @ gmail.com
@inproceedings{zaheer2020old,
title={Old is Gold: Redefining the Adversarially Learned One-Class Classifier Training Paradigm},
author={Zaheer, Muhammad Zaigham and Lee, Jin-ha and Astrid, Marcella and Lee, Seung-Ik},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={14183--14193},
year={2020}
}