search

zhiyuanyou / UniAD

[NeurIPS 2022 Spotlight] A Unified Model for Multi-class Anomaly Detection
Apache License 2.0
250 stars 28 forks source link

How to get 128x128 cifar10 images? #3

Closed coordxyz closed 2 years ago

coordxyz commented 2 years ago

If I simply resize the image with bicubic interplotation, the resized images will contain a lot of griding(chessboard) noise. I also find a 128x128 version of cifar10 by using super resolution tech from https://www.kaggle.com/datasets/joaopauloschuler/cifar10-128x128-resized-via-cai-super-resolution. But the SR images also contain lots of other kinds of artifacts. I guess all these artifacts will effect the AD performance. Do you have similar problem? How do you get the 128x128 size for cifar10 images whose original size is 32x32? Thx~

zhiyuanyou commented 2 years ago

We are now working to provide the CIFAR-10 results. The code would be released in one week.

About how to resize, you could see Line103 of https://github.com/zhiyuanyou/UniAD/blob/main/datasets/cifar_dataset.py, i.e., transforms.Resize(resize, Image.ANTIALIAS).

coordxyz commented 2 years ago

Thanks, I'll try to use Image.ANTIALIAS rather than bicubic.

zhiyuanyou commented 2 years ago

We have added the CIFAR-10 results. If you still have questions, please re-open this issue.