Great work! I read this paper and just wonder if the idea is similar to the classic 10-crop testing used in the ImageNet benchmark in the "old days" (like below, cropped from the AlexNet paper).
Here, a few learnable transformation layers are trained vs predefined transformations used in the past 10-crop testing. Essentially, it is like a kind of (learnable) data augmentation? Open to discussions. Thanks!
Regards,
Great work! I read this paper and just wonder if the idea is similar to the classic 10-crop testing used in the ImageNet benchmark in the "old days" (like below, cropped from the AlexNet paper).
Here, a few learnable transformation layers are trained vs predefined transformations used in the past 10-crop testing. Essentially, it is like a kind of (learnable) data augmentation? Open to discussions. Thanks!
Regards,