Loss ratio and image alignment

[llp1996]

could you tell me the ratio of loss on diff level of pyramid?

i find the code caculate the loss of each level of laplace pyramid .is the loss ratio 1:1:1? i train the model with a ratio of 1:1:1,but i find that the network focal on minimizing the loss on the last level (Ln). so the model cant perform well (netout is almost same as input) . i change it to 100: 100 : 1,and model performs better.

thanks

[csjcai]

@llp1996 To compare with other methods in quantitative metrics (calculate the PSNR and SSIM), we set the loss ratio to 1:1:1. While for visual quality, we set the loss ratio to 16:4:1.

[llp1996]

i run the align code with my own data ,the result is not as good as dataset provided. I process the HR and LR image of same resolution without any pre-procession. is it right? thanks

[csjcai]

What kind of data you used? In this work, we first make sure the collected image pairs do not have a significant misalignment. That means, at the dataset collection stage, you should take a lot of time to capture image pairs. After the collection stage, we remove the lens distortion for both LR and HR, and then we use the provided alignment code. You should adjust the parameter for different scales. It is a non-trivial work to construct this kind of dataset.

[csjcai]

At the alignment stage, we (1) coarse align the image pairs; (2) central crop the image pairs; (3) finer align the cropped image pairs; (4) discard those misaligned image pairs.

[llp1996]

thanks for align details