Datasets:

maelic
/

Haystack-coco-format

license: other
license_name: sa-1b-dataset-research-license
license_link: >-
  https://huggingface.co/datasets/maelic/Haystack-coco-format/blob/main/LICENSE_SA-1B
task_categories:
  - object-detection
tags:
  - scene-graph-generation
  - panoptic-scene-graph
  - negative-annotations
  - haystack
  - sa-1b
  - coco-format
language:
  - en
pretty_name: Haystack (COCO format)
size_categories:
  - 10K<n<100K

Haystack (COCO format)

A mirror of Haystack (Lorenz, Barthel, Kienzle and Lienhart, Haystack: A Panoptic Scene Graph Dataset to Evaluate Rare Predicate Classes, ICCV 2023 workshops), reformatted in the same COCO-style schema as the other maelic/*-coco-format datasets. Haystack's distinguishing feature is explicit negative relation annotations: for an annotated subject–object pair, an annotator also confirmed predicates that do not hold, so evaluation does not have to treat missing labels as negatives.

It is used by axis A2 (precision against adjudicated negatives) of OV-SGG-Bench.

Why this mirror exists

The original download location (myweb.rz.uni-augsburg.de/~lorenjul/haystack/v1/) returns HTTP 404 as of November 2026. This repository preserves the v1 release byte-for-byte where possible so that results on Haystack remain reproducible. The authors' code is at https://github.com/lorjul/haystack.

Licenses

Images are a subset of Meta's SA-1B dataset and are provided under the SA-1B Dataset Research License (LICENSE_SA-1B, with Meta's notice in NOTICE_SA-1B.md): research, non-commercial use only. By downloading the images you agree to that license. See https://ai.meta.com/datasets/segment-anything/.
Annotations (relations, negative relations, panoptic masks) are the Haystack authors' work. Their release did not state a license. Cite their paper when you use them.

Samples

Four images drawn at random (fixed seed) from the 61 that have at least two positive relations, two to eight negatives and at most five objects in relations, with the boxes of every object that takes part in an annotated relation. Green lines are positive relations; red lines are negatives an annotator explicitly rejected for the same pair.

One split, test, with 11,368 images, 2,870 positive relations and 23,174 negative relations over 133 categories (80 things, 53 stuff) and the 56 predicates of PSG. categories.json holds thing_classes, stuff_classes, categories (things then stuff, the index used by category_id) and predicates (the index used by predicate_id).

Field	Description
`image`	the SA-1B JPEG, original bytes
`image_id`, `width`, `height`, `file_name`	as in the source (`sa_XXXXXX/sa_N.jpg`)
`pan_seg`, `pan_seg_file_name`	the panoptic segmentation PNG, original bytes
`objects`	`id` (index in the image), `bbox` (COCO `[x, y, w, h]`, pixels), `category_id`, `area` (segment area), `iscrowd`, `isthing`, `segment_id` (the segment's id in `pan_seg`)
`relations`	positive relations: `id`, `subject_id`, `object_id` (indices into `objects`), `predicate_id`
`neg_relations`	adjudicated negative relations, same fields

Changes from the source

All mechanical: boxes converted from Detectron2 XYXY_ABS to COCO [x, y, w, h]; object and relation ids made image-local indices; images and masks stored with their original bytes, not re-encoded. A random sample of 50 images was checked byte-for-byte against the source archives. Source archive SHA-256: sa1b-subset.zip cff441c94ba7ce3af5b000a5b1648e9610e3357241338698e3bd79cb396bb919, v1.zip 07401be66aa4f048a3d52e12b624f393166654b42a4275cb5a051de7fe7ebee4.

Citation

@inproceedings{lorenz2023haystack,
  title     = {Haystack: A Panoptic Scene Graph Dataset to Evaluate Rare Predicate Classes},
  author    = {Lorenz, Julian and Barthel, Florian and Kienzle, Daniel and Lienhart, Rainer},
  booktitle = {IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)},
  pages     = {62--70},
  year      = {2023},
  doi       = {10.1109/ICCVW60793.2023.00013}
}