search

varun19299 / deep-atrous-guided-filter

Deep Atrous Guided Filter for Image Restoration in Under Display Cameras (UDC Challenge, ECCV 2020).
https://varun19299.github.io/deep-atrous-guided-filter/
36 stars 6 forks source link

Can we use 128*128*3 patch to train and 1024*2048*3 image to test ? #2

Closed hellolingling closed 3 years ago

hellolingling commented 3 years ago

Hi @varun19299, can we use 1281283 patch to train and 102420483 image to test ? When I use 1281283 patch to train a new model and inference a image with 102420483, I found the result is not correct. But if I split the image into multiple 1281283 path to inference, the result seems much better. How can I change your code to overcome the problem? waiting for your help!

varun19299 commented 3 years ago

Yes, this is something we talk about in our paper. Our architecture is designed to directly train on 2MP images, which attains superior results, especially on POLED measurements.