Giters

Chintan2108 / Cloud-Removal-in-Satellite-Images-using-Conditional-Generative-Adversarial-Networks

Removing cloud cover in Sentinel-2 satellite images using only optical data and a novel augmented training approach using conditional GANs.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Implement skew for dataset augmentation

Chintan2108 opened this issue · comments

commented
  • Currently the training dataset is augmented from a single pair of cloudy and cloud-free images
  • The augmentation is performed only by rotation of 0, 90, 180 and -90 degrees which gives us 4 pairs, say 1,2,3 and 4
  • These four pairs are then stacked iteratively 10 times as: [1,2,3,4,1,2,3,4,1,2]

  • Implement skew operations on the cloudy and cloud-free images, and then perform the rotation operations. You will now have 4 pairs of original (unskewed) images, say, [1, 2, 3, 4] and 4 pairs of the skewed images, say [1', 2', 3', 4']

  • Prepare two new training datasets by stacking 10 times as [1,2,3,4,1',2',3',4',1,1']

commented

Hey, @Chintan2108 you would want a Jupyter Notebook for the Data Augmentation on the images or a Python file.

commented

Hey, @Chintan2108 you would want a Jupyter Notebook for the Data Augmentation on the images or a Python file.

Hi @Preetesh21 Ideally it would be great if you could add this logic in the sections 3.3 and 3.4 of the current notebook.
Even if you do it in a separate file, it's okay. Make sure you write your code with proper docstrings and comments. Just make sure you upload the new dataset in the drive folder with appropriate naming convention.