metadata
license: cc-by-4.0
size_categories:
- 10K<n<100K
SharvitSR is a synthetic dataset for spatial relation reasoning. The dataset contains 44100 images, each containing two objects (one green, one yellow).
To use the dataset, download both the .zip file (which contains the images) and the .csv annotation file. SharvitSR_annotations.csv contains six columns:
- 'obj1' and 'obj2' refer to the name of the image related to the row ('img-{obj1}-{obj2}.png')
- 'card' contains the directional relation between the two objects.
- 'dist' refers to the distance between the two objects.
- 'depth' refers to the overlap of one object over another.
- 'diff' refers to the disagreement of models over the directional relation.
directional relations: [Above, Above Right, Right, Under Right, Under, Under Left, Left, Above Left].
distance relations: [Near, Mid, Far].
Depth relations: [Front, Behind, None].
If you use our dataset, please cite our work:
@article{SERVANT2026112013,
title = {SpIRL: Spatially-aware image representation learning under the supervision of relative position descriptors},
author = {Logan Servant and Michaël Clément and Laurent Wendling and Camille Kurtz}
journal = {Pattern Recognition},
volume = {170},
pages = {112013},
year = {2026},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2025.112013},
url = {https://www.sciencedirect.com/science/article/pii/S0031320325006739}
}