I have a custom database made of bw superficial defects images. The database is quite far from classical CV Dataset, like CIFAR or ImageNet. I know form supervised Deep Learning that the correct choice of hyperparameters can have a dramatic effect on the model performances. In the image classifier that I developed, the correct number of epochs and the appropriate learning lead to a 15% increase of the test set accuracy.
What are the target metrics for GANs. I know and use multi-scale version of SSIM, the PNSR, and the generator/discriminator losses. But from experiments I see no correlation with the quality of images by human eye inspection. I have also many doubts on metrics like IS and FID, because they rely on pre-trained models, often on ImageNet weights.
I am also aware that this is a recent and open recent topics, but I would like to get some tips and opinions.
