search

Shaosifan / HSENet

Hybrid-Scale Self-Similarity Exploitation for Remote Sensing Image Super-Resolution (accepted by TGRS)
37 stars 5 forks source link

Questions on training datasets #11

Closed xiaying-bot closed 3 months ago

xiaying-bot commented 4 months ago

hi,

I noticed in your paper, the dataset is split into two balanced halves as training and test sets with 1050 samples each, however the datasets I downloaded from the provided link has 1050 test samples, 946 train samples and 105 val samples.

I am confused and wonder which setting did you exactly use in your experiment

Very much kindly looking forward to your reply!

Shaosifan commented 4 months ago

hi,

I noticed in your paper, the dataset is split into two balanced halves as training and test sets with 1050 samples each, however the datasets I downloaded from the provided link has 1050 test samples, 946 train samples and 105 val samples.

I am confused and wonder which setting did you exactly use in your experiment

Very much kindly looking forward to your reply!

Hi, in our paper, the dataset is split into two balanced halves including training and test sets. To select the model in the training phase, we further split the training set into the training subset with 945 samples and the val subset with 105 samples.

It should be noticed that this implementation is a little different from the setting of the natural image super-resolution where the test sets such as Set5 and Set 14 also be taken as the val sets in the training phase.