Hello guys!
I have a small question. My images have a size of 224x224 pixels. But when I looked into the code of pytorch’s Faster R-CNN, it was mentioned that all images will be rescaled in the code. I think, they used GeneralizedRCNNTransform for this purpose. So when I select the anchor sizes, should I only consider the sizes of my annotations on the 224x224 original image? Or should I adapt the anchor sizes to the sizes of my annotation in the transformed image?
Would be happy to get an answer.