Inquiry Regarding Image Scaling and Cropping Methodology

[thomashirtz]

Hello Neural-ILT team,

I've been delving into the Neural-ILT project and am quite impressed with the work you've achieved. While examining the image processing steps taken prior to inputting images into the network, I noted a distinct approach to scaling and cropping. This approach piqued my interest, as it appears to diverge from traditional methods that typically involve resizing images to a uniform dimension to preserve scale, followed by a process of random cropping.

Could you please provide some insight into the rationale behind this preprocessing strategy? Specifically, I'm curious about the reasons for saving different values for each image and applying custom cropping. Is this approach essential? Additionally, considering there's scaling involved, could you inform me about the image scale at which the kernels are designed to operate, especially if I'm considering doing custom cropping?

Thank you for your time and consideration. I eagerly await your response :)

Thomas,