readme.md

Data

If you want to use your own data, you can copy it into these folders or mount the folders correctly if you use Docker ( see docker_run.sh for details)

Note, that you need a split.json file, with the following form (see e.g. split.json)

{
  "test": [
  ],
  "train": [
    "misc/transport-ga1a289b15_1920.png",
    "misc/clock-g204ee03ae_1920.png",
    "misc/kerosene-lamp-gde06bdaab_1920.png",
    "misc/teddy-bear-g08d67e00b_1920.png",
    "misc/vw-gebf7ea815_1920.png",
    "misc/chair-g1b12b7580_1920.png"
  ],
  "validation": [
  ]
}

• it contains a list for each split (note, that for this mini demo, we only have train data)
• within each list, we have relative paths to the images
• optionally it can contain a fourth key path with the base path of the images (if they are not inside this folder)

Background

• Images were downloaded from Pixabay, see sources.txt for the links.
• For our dataset we used images from SUN Database: Large-scale Scene Recognition from Abbey to Zoo
  • download complete dataset with wget "http://vision.princeton.edu/projects/2010/SUN/SUN397.tar.gz"
  • unzip with tar -xf SUN397.tar.gz

Distractors

Images were downloaded from Pixabay, see sources.txt for the links. Afterwards, we applied tight-cropping (see tight_crop.py).

Objects

Images are photos taken by myself, where the background was removed using rembg Afterwards, we applied tight-cropping (see tight_crop.py).