I have a pre-trained unet segmentation model trained on a particular image dataset. The model can not segment well on a new dataset that comes from the same domain but differs in quality.
But when I set the moving_mean=None and moving_var=None in the batch norm layer of the model, it starts to give good segmentation results on this new dataset in evaluation/inference mode.
What does this mean? Can we use this modification of the model during deployment?