Question about the CIFAR dataset about eccv22-foster HOT 2 CLOSED

Thanks for your help.

from eccv22-foster.

Thank you so much for your interest in our work.

DER claims they use the standard resnet-18 rather than the modified resnet-32 for CIFAR-100, and the following is copied from the implementation details in the paper.

We note that most previous works use a modified 32-layers ResNet, which has fewer channels and residual blocks compared to standard ResNet-32. We argue that such a small network is not suitable because it cannot achieve competitive results on CIFAR100 compared with standard 18-layers ResNet and may underestimate the performance of methods. We run these methods with standard ResNet-18 on the same class orders based on their code implementation.

They also report the performance on CIFAR-100 with the modified 32-layers ResNet in the appendices (table 7, 8), and you can refer to that for a fair comparison.

Hope this clarifies your queries.

from eccv22-foster.