Given an old model and you have a new model, and both models return the prob for each pixel that it is a human or not, how would you determine the quality of the two models? What if the test data is very hard or very easy, how would that change your test result? (IoU, Dice coefficient, Pixel accuracy)