Will fine-tune with cropped image influence the performance

vision
1

Hi there,

Currently, I have a well-trained segmentation model on GTA5. I’d like to further fine-tune this trained model on a self-defined cropped gta5 image. Now I have two choices:

  1. crop the original image and input a small-sized cropped version into the network.
  2. keep the image as original size and replace all non-crop regions with 0 (non-cropped region will be labeled as 255).

I was just wondering if these two operations have significantly different performance.

Thank you!