The format of submission is a single .zip-file that should include 290 super-resolution images, one for each data member in the test-set.
Each submitted image must be in .png-format, saved in 16bit grayscale in a 384x384 pixel shape. The file should be named "imgsetxxxx.png" where xxxx corresponds to the id of the test-set (case sensitive).
Validation in detail
The validation step performs the following checks in that order:
size of file is less than 100MB
file needs to have the extension ".zip"
file needs to be a valid zip-file (no corrupt header, no other compression format allowed)
none of the 290 images are missing
there are no additional files in the zip except the 290 images
the images in the zip are named from "imgset1160.png" to "imgset1449.png"
every image is 16bit grayscale png
every image has a resolution of 384x384 pixels
If one of the checks fails, you will be notified with a corresponding error message. If your submission is valid, it will be evaluated and be shown on the public leader-board as soon as its score) has been computed (might take a minute).
You can download the following python script which will generate an example submission for you.
Assuming that all data is available at a local folder data, run the script with:
python generate_sample_submission.py data submission
If everything works out, it will create a folder submission, containing some images and a .zip-file which you can use to test the system.
The algorithm in this script takes all low resolution images that have the maximum amount of clear pixels, computes a bicubic upscaling by a factor of 3 and averages their pixel intensities. The Scoring of these images (in terms of their \(cPSNR\)) constitutes our baseline of value 1 for ranking submissions. Thus, in order to improve upon bicubic upscaling, a submission needs to receive a score lower than 1.