Question about fully convolution prediction and sliding window


I’m doing a task of semantic segmentation:
I have trained a fully-convolutional CNN (like Tiramisu) using a segment_generator() on random slices of 512,512 (for example) without resize.
I would now like to do model.predict on the entire image and considering two approaches:

  1. Since this is a fully convolutional CNN - I can input the entire image (let’s say 1920x1280) and get a prediction
  2. Use a sliding window loop

For some reason, when using a sliding window I am getting significantly better results (dice 0.996 vs 0.993).

Any ideas why I’m experiencing this?