search

hq-deng / RD4AD

Anomaly Detection via Reverse Distillation from One-Class Embedding
MIT License
174 stars 31 forks source link

cifar #7

Closed xingp-ng closed 2 years ago

xingp-ng commented 2 years ago

Regarding the CIFAR and MNIST datasets if they are 3232 in size, the 77 convolution in ResNet should not apply, did the authors make a change?

hq-deng commented 2 years ago

I didn't change it. In fact, the current method does not work well on a dataset like Cifar10. As we use the pre-trained model on ImageNet, the adaptability of the model is a drawback. Our model is assumed to work well on natural images similar to ImageNet. To make the model more adaptive, you can use method like CutPaste[1] or DREAM[2] to re-train the pre-trained model and you can also modify the architecture.

[1]Li, Chun-Liang, et al. "Cutpaste: Self-supervised learning for anomaly detection and localization." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. [2]Zavrtanik, Vitjan, Matej Kristan, and Danijel Skočaj. "DRAEM-A discriminatively trained reconstruction embedding for surface anomaly detection." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021.