SPARE-Corpus

This repository holds the Birmingham-Tufts Spatial Referring Expression Corpus.

Organization

• corpus/ contains data.csv which has four columns:
  • Description: an RE collected during our experiment.
  • image_id: the image shown to the participant during the experiment.
  • scene_id: the scene configuration used to generate that image.
  • Object Type: the type for the target referent.
• img/ contains a set of .png files named with corresponding image_ids.
• metadata/ contains scenes.json, which contains a list of image metadata entries. Each entry maps an imageid to a scene description, which contains:
  • pixel: a list of entries that are either:
    • WIDTH: the width of the image
    • HEIGHT: the height of the image
    • or an object name mapped to its bounding box within the scene
  • object: the name of the object bounded within the scene
  • scene_id: the scene_id for this configuration.

License

This corpus is released under a Creative Commons Attribution-ShareAlike 4.0 International license.

If you use this data, please cite "Spatial Referring Expression Generation for HRI: Algorithms and Evaluation Framework" by Lars Kunze, Tom Williams, Nick Hawes, and Matthias Scheutz, as appeared in the proceedings of the 2017 AAAI Fall Symposium on Artificial Intelligence for Human-Robot Interaction (AR-HRI).

   @InProceedings{kunze2017arhri,
        author = {Lars Kunze and Tom Williams and Nick Hawes and Matthias Scheutz},
        title = {Spatial Referring Expression Generation for HRI: Algorithms and Evaluation Framework},
        booktitle = {{AAAI} Fall Symposium on Artificial Intelligence for Human-Robot Interaction ({AR-HRI})},
        year = {2017},
        address = {Arlington, VA, US}
        }

Note that that paper used a subset of this corpus corresponding to data collected from participants that faithfully completed the entire experiment.