maelic/SpatialSense-coco-format

SpatialSense (COCO format)

A mirror of SpatialSense (Yang, Russakovsky and Deng, SpatialSense: An Adversarially Crowdsourced Benchmark for Spatial Relation Recognition, ICCV 2019), reformatted in the same COCO-style schema as the other maelic/*-coco-format datasets. Every annotation states whether a spatial predicate holds between a subject and an object, and the dataset is balanced: half the relations are true and half are false.

It is used by axis A6 (spatial relations) of OV-SGG-Bench.

Source and license

The authors' release is Zenodo record 8104370 (DOI 10.5281/zenodo.8104370), published under CC BY 2.0; this mirror keeps that license and attributes the authors. Images come from Flickr and the NYU Depth dataset, as in the original release. The two source files were checked against the SHA-256 sums the record publishes (annotations.json ae8f613ff1886e27e06f154cc01393c8a86025ae7f8f3664a7fd2308d771bf39, images.tar.gz 1f9db2e56f323046aec301288a7e24bd35cce51f651cab93ea2e41dc27b853bd).

Samples

Four test images drawn at random (fixed seed) from the 22 that have at least three relations, both labels and at most five objects, with the boxes of every annotated subject and object; an object annotated twice with slightly different boxes appears twice. Green lines are relations labelled true, red lines relations labelled false.

Contents

Split	Images
train	7,645
validation	1,553
test	2,371

11,569 images, 8,749 true and 8,749 false relations, 9 predicates (above, behind, in, in front of, next to, on, to the left of, to the right of, under) and 3,679 free-text object names. categories.json holds categories (the index used by category_id) and predicates (the index used by predicate_id), both sorted alphabetically.

Field	Description
image	original bytes (Flickr JPEG or NYU PNG)
image_id	index of the image in the source annotations.json
width, height, file_name, url, nsid	as in the source; file_name is the path inside images.tar.gz
objects	subjects and objects deduplicated by (name, box): id, bbox (COCO [x, y, w, h], pixels), category_id, area, iscrowd
relations	relations labelled true: id, annotation_id (the source _id), subject_id, object_id (indices into objects), predicate_id, subject_point, object_point (the [x, y] points the annotator clicked)
neg_relations	relations labelled false, same fields

Changes from the source

All mechanical: boxes converted from SpatialSense's [y_min, y_max, x_min, x_max] to COCO [x, y, w, h]; objects deduplicated per image; the label field turned into the relations / neg_relations split; macOS ._* files from the archive skipped; images stored with their original bytes. A random sample of 50 images was checked against the source.

Citation

@inproceedings{yang2019spatialsense,
  title     = {SpatialSense: An Adversarially Crowdsourced Benchmark for Spatial Relation Recognition},
  author    = {Yang, Kaiyu and Russakovsky, Olga and Deng, Jia},
  booktitle = {IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages     = {2051--2060},
  year      = {2019},
  doi       = {10.1109/ICCV.2019.00214}
}