Skip to content

Preparing data for the MIT Adobe FiveK Dataset with Lightroom

Jump to bottom
Yuanming Hu edited this page Jul 9, 2018 · 4 revisions

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 (~10 minutes).

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. Resolution doesn't matter since it is not the actual image resolution in pixels.
    • Finally, click Export.
  • You can compare the first exported image (0001.tif) with this image. If you have done the previous steps correctly, you should get an identical image. Or, you can simply download the exported image pack Part I Part II.
  • Execute fivek.py and you will get the augmented image pack at exposure/data/fivek_dataset/sup_batched80aug_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. Resolution doesn't matter since it is not the actual image resolution in pixels.
    • Finally, click Export.
  • You can compare the first exported image (0001.jpg) with this image. If you have done the previous steps correctly, you should get an identical image.
Clone this wiki locally