Below we describe the contents of the OpenMIC files.

1. Folder structure

1.1. Folders with prefix 'crops' such as 'crops_clk, crops_gls, crops_nat, crops_sci, crops_shn, crops_clv, crops_hon, crops_rel, crops_scl, crops_shx,' contain images of three crops extracted from each original image (left, center, right, or top, middle, bottom depending on aspect rato). These are the crops used in our experiment in ECCV'18 paper and few-shot learnign paper.

Folders 'crops_source' contain source iamges taken with mobile phones from different viewpoints and distances. They typically contain only a single exhibit with single label.

Folders 'crops_target' contain target iamges taken with our wearable cameras. They often contain multiple exhibits.

1.2. Folders with prefix 'full' contain the original images so these are not square region images. It contains 'full_source' and 'full_target' folders.


2. Labelling

2.1. Files below contain all labels (no data splits):

crops_source.txt : contains all source labes (from 0 to C)
full_source.txt  : contains all source labes

crops_target.txt : contains all target labes (only the most sailient label is listed per image)
full_target.txt  : contains all target labes (only the most sailient label is listed per image)

2.2. The following folder contains our ECCV'18 target data splits (for the source we use the files described in Section 2.1):

Folder 'target_splits_eccv2018' contais target splits used by us in our ECCV'18 paper.

The following files contain the training, validation and testing splits (5 splits) with only the most salient label assigned to each image.
full_train_3_sp[1..5].txt
full_val_3_sp[1..5].txt
full_test_3_sp[1..5].txt

If you train on split k, use validation on k and testing on k too, then compute average and the standard deviation to report your results.

The following files contain the training, validation and testing splits (5 splits) with all labels assigned to each image according to the saliency order (the first label corresponds to the most central/significant object followed by less and less significant objects):
full_train_3_sp[1..5]_multilabelfix.txt
full_val_3_sp[1..5]_multilabelfix.txt
full_test_3_sp[1..5]_multilabelfix.txt

The 'multilabelfix' splits contain the exactly same files as 'single label' splits.

The 'full' and 'crops' prefixes denote versions for full images and crops, respectively, and they contain the exactly same files in the same order but, obviously, the 'crops' versions have 3 crops per image (hence 3x longer list).

2.3. Folder 'multilabels_target' contains complete source and target lists in 'bgr_source.txt' and 'bgr_target.txt' which are the same as files in Section 2.1, however, they additionally contain background labels '-1' which are not exhibits. File 'bgr_target_multilabels.txt' additionally contain full target lists with all labels assigned to each image according to the saliency order, as in Section 2.2, as well as background labels denoted as '-1'.

2.4. Folder 'distortion_attributes' contains annotations for the geometric and photometric distortions observed in target images. We used these annotations in our ECCV'18 paper.


3. Different resolutions:

The zip files with prefixes '256_' and '512_' contain images whose smaller dimension has been resized to either 256 or 512px resolution. This way, the dataset is smaller and easier for downlaod. Note that the cropped versions contain exactly 256x256 or 512x512px images.

4. If using images for illustrating your work in papers, please follow the license file. Remember also not to use any images that contain humans in it (if one can see their face) or iamges that may be copyrighted by third parties. The rule of thumb is that exhibits which are very old are leass likely to be copyrighted (but this does not have to be always true).

Lastly, check '_license.txt' file and remember to adhere to its rules if you accepted it. In any cases, contact us via e-mail provided on the page. Also, if you find that for some reason you would like to make some modifications to our dataset, please check our 'no derivative' clause. In such cases we are happy to talk to you about how to accomodate your changes.