andrea-covre
commented
1 year ago I ran into the same issue as well while creating my own dataset, saving the images in PNG format is a much better solution that won't generate compression artifacts. If you are looking for the CIFAR-10 dataset with the raw images you can find it here https://www.kaggle.com/competitions/cifar-10/data
Hello, We used this version of CIFAR10 database to train models in timm (https://github.com/rwightman/pytorch-image-models), and it lead systematically to a drop of 2-3% accuracy. Upon inspection, it looks like this version of the dataset has substantial JPEG artifacts -- the images actually look noticeably less sharp and colorful. This is the topic of this issue that you can read here: (https://github.com/tmp-iclr/convmixer/issues/11#issuecomment-951622097).