Informative Object Annotations (IOTA) - CVPR 2019

Code for selecting informative labels of images based on a known distribution of labels.

Paper

See our CVPR 2019 Paper

Citation

If you find IOTA useful in your research, please consider citing:

@inproceedings{bracha2019informative,
  title={Informative Object Annotations: Tell Me Something I Don't Know},
  author={Bracha, Lior and Chechik, Gal},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={12507--12515},
  year={2019}
}

Download code

git clone ssh://git@github.com:liorbracha/iota.git

To install the anaconda environment

$ conda install --yes --file requirements.txt

to install in a “single package mode”:
$ while read requirement; do conda install --yes $requirement; done < requirements.txt

IOTA-10K ground truth data

wget https://chechiklab.biu.ac.il/~brachalior/IOTA/data/iota10K/iota_raw.csv.tar.gz

Open Images Dataset (OID)

Download Image-Level Labels from the Open Images Dataset to data/oid

To change the default location of the data or results directory:

export RES_DIR=/where/results/are/written
export OID_DIR=/where/data/is/located/

# data
#   |_ ground_truth
#   |_oid
#       |_classes
#       |_train
#       |_validation
#       |_test
# results
#   |_models
#   |_counts