Preparing data for the MIT Adobe FiveK Dataset with Lightroom

Getting the data

• Download the dataset from https://data.csail.mit.edu/graphics/fivek/. (The "single archive (~50GB, SHA1)" or "by parts", either is fine)
• Extract the data
• Open the file fivek.lrcat. Lightroom may probably ask you to upgrade. Just click "upgrade" if you are asked to. You may need to wait for a while.

Generating the Training Input Set

• Open the dataset with Adobe Lightroom
• In the Collections list, select collection 'Inputs/Input with Daylight WhiteBalance minus 1.5'
• Select all images in the bottom (select one and press Ctrl-A), right-click on any of them, choose Export/Export...
  • Export Location: Export to=Specific folder, Folder=exposure/data/fivek_dataset/FiveK_Lightroom_Export_InputDayLight/
  • Image Format=TIFF, Bit Depth=16 bit/component (input images have to be RAW). Compression=None. Color Space=ProPhoto RGB
  • Image Sizing: Resize to Fit=Long Edge. Click Don't Enlarge. Fill in 500 pixels
  • Finally, click Export
• You can compare the exported first image (0001.tif) with this image. If you have done the previous steps correctly, you should get an identical image.
• Execute fivek.py and you will get the augmented image pack at exposure/data/fivek_dataset/sup_batched%daug_daylight in minutes.

Generating the Training Target Set (Expert C)

• Open the dataset with Adobe Lightroom
• In the Collections list, select collection 'Experts/C'
• Select all images in the bottom (select one and press Ctrl-A), right-click on any of them, choose Export/Export...
  • Export Location: Export to=Specific folder, Folder=exposure/data/artists/Five_K
  • Image Format=JPEG (Target images do not have to be RAW so using jpg should be fine). Quality=92. Color Space=sRGB
  • Image Sizing: Resize to Fit=Long Edge. Click Don't Enlarge. Fill in 500 pixels
  • Finally, click Export
• You can compare the exported first image (0001.jpg) with this image. If you have done the previous steps correctly, you should get an identical image.