Closed circlehy closed 4 years ago
Besides, the code seems used different parameter for losses from the paper paper: code parameters: Please correct me if I configed the code wrong.
Q1: We assume the GT kernel is unknown and also its size. KernelGAN estimates a 13x13, 25x25 kernel for X2, X4 respectively. These are sizes we found give a wide enough and "natural" support (a larger one blurs too much and a smaller is not expressive enough). If you want to compare two different size kernels - you may pad the smaller one with zeros without effecting its blur function.
Q2: As written in the paper, all the regularizations are inserted after the bicubic constraint is satisfied and discarded. The final co-efficients are as written in the paper (and you quoted). The parameters you snapped are at the beginning of the training before the bicubic "satisfaction".
Training with image 'DIV2KRK/lr_x2/im_78.png', the ground truth kernel should be 'gt_k_x2/kernel_78.mat' right? The generated result kernel size seems different from the ground truth kernel? Is this the final result?