ResNet input requirements. Segmentation

Hi everyone!

I am working on the task of semantic segmentation and my choice of the model backbone is ResNet. Specifically, I use ResNet-34 with imagenet prettained weights.

However, it is not clear what input requirements the model has. The only thing I found is 224x224 dimensions. But what processing should be done? Should it be normalized to 0…1 or -1…+1 range? How important to maintain mean and std values of channels? Maybe some other things to be done? How do you find such information in general and how important it is to has exactly these values?

I’ve seen tensorflow has preprocessing function, but in my case it puts values to random -200 values, I think for my image input it may be doing something wrong.

Thank you for your time and help!

You should check the resnet relevant model in tensorflow about instructions (starting searching on google). Normally they include a preprocessing function and also try to find the github repo if you need it to find details about it. As far as I know they have a prerpocessing function that does it for the input!