Gradient metric different from reference paper

[poier]

Hi,

first, very nice work. Thanks.

In the paper you refer to [Rhemann et al., CVPR 2009] for the gradient based (GRAD) metric. In [Rhemann et al., CVPR 2009] they compute this as the sum of the distance between the normalized gradients of the prediction and ground truth. However, in you implementation you compute the distance between the gradient magnitudes (and divide the sum of the squared distance by 10; https://github.com/JizhiziLi/P3M/blob/master/core/evaluate.py#L154-L158). What is the reason for that?

Best,
Georg

[Rhemann et al., CVPR 2009] https://www.microsoft.com/en-us/research/wp-content/uploads/2009/01/cvpr09-matting-Eval_TR.pdf