I am wondering if there is any existing generations of ResNet (or any architecture) taking large image inputs, such as 1000 by 1000 without resizing to 224 by 224. I have a few high resolution images and fairly sufficient GPU resources, so I want to train the model on them directly. Please share some thoughts. Thanks in advance.
The current ResNet implementaions are flexible regarding the spatial input size (as long as it’s not too small) and would accept the 1000x1000 input images (assuming your GPU has enough memory).
I’m not aware of any pretrained models using this resolution so you might want to train the model from scratch (or fine-tune it).