Jamy-L / Handheld-Multi-Frame-Super-Resolution

Handheld Multi-image Super-resolution [Wronski et al., SIGGRAPH19]. Non-official GPU-supported Python implementation.
MIT License
117 stars 20 forks source link

when set 'kernel_shape=iso', run with cuda error, don't support this param? #24

Closed renwoxing000 closed 1 year ago

Jamy-L commented 1 year ago

Hi, after checking on my side on the latest version, I had no such problem. Make sure you are specifying the shape like this :

params={
        "scale":2,
        "merging" : {
            'kernel': 'iso'},
        }

If the problem persists, please use a verbose of level 4 using :

options = {'verbose': 4}

and print what you get.

Best, Jamy

renwoxing000 commented 1 year ago

Hi Jamy:

I check the demo with param 'kernel = iso' in newest code, there is no problem.

By the way, the param transmission has some problem. when config params like this: params={ "scale":2, "merging" : { 'kernel': 'iso'}, }

Param "kernel iso " not valid in the function "merge()". The reason of no valid is the params of 'kernel' is overwrite by param init process.

renwoxing000 commented 1 year ago

Hi Jamy:

The results of isotropic kernel and anisotropic kernel looks like same ?

Jamy-L commented 1 year ago

Hi, after checking everything on my side, I have found no issue with the transmission of the parameters. the function merge() takes the params dictionnary as input and reads it as expected. For the results, the difference has been illustrated in the article and can be observed by zooming agressively : isotropic kernels make edges more blurry and are not as good as steerable kernels for denoising

renwoxing000 commented 1 year ago

I will check again, thanks.