About experiments setting
Opened this issue · 3 comments
Hello. I found your SCPNet paper very impressive and enjoyed reading it. I have 3 questions regarding the experiments:
-
In the paper, you mentioned, 'For the RGB/NIR dataset, we use 103 pairs of images for training and 153 pairs for testing.' Could you provide more details on how these pairs were selected?(or which pairs were selected?)
-
For 'flash' & 'harvard' dataset, how to divide train/test? Which pairs were selected to be training or testing?
-
In the provided code, it seems that A.RandomBrightnessContrast() is applied each time getitem() is called on the dataset. However, I didn't see this mentioned in the paper. Could you please confirm if this is correct?
Thank you in advance for your assistance.
Thanks for your interest.
- (Q1 & Q2) The training/test pairs are randomly divided. We have uploaded the supplementary material in this repo (docs/SCPNet-supp.pdf). Please refer to it for more details.
- (Q3) A.RandomBrightnessContrast() is a common data augmentation technique. We use it in intra-modal self-supervised learning to further enhance the generalization ability.
Hello, is the result obtained by your reproduction accurate? I tried to reproduce the paper with ggmap, but the result was not satisfactory.
Hello, is the result obtained by your reproduction accurate? I tried to reproduce the paper with ggmap, but the result was not satisfactory.
Hello, here is printed log.
"
start testing
|Test size | mace12 | mace11 | mace22 |
111 4.9457870732556595 1.1636654444642969 2.4171450089763953
end testing
"