search

goutamgmb / NTIRE21_BURSTSR

85 stars 15 forks source link

About the scale factor in real-world track #1

Closed NiPaung closed 3 years ago

NiPaung commented 3 years ago

hello, i wonder the scale factor of the real-world track.

The description section of this github says that the goal is to predict 4 times higher resolution image. However, In train and validation datasets, HR images 8 times larger than LR images and python code is written to set scale factor with x8.

Is it correct to set the scale as provided in the python toolkit?

goutamgmb commented 3 years ago

Hi,

Given the s x s RAW bayer mosaic as input, the task is to predict a 4s x 4s RGB image. However, we pack the RAW image as a 4 channel tensor s/2 x s/2 x 4, where the channels correspond to the R, G, G, B values. Hence there is an additional 2x factor to perform demosaicking.

Regards