Haystack v1

Dataset for our paper Haystack: A Panoptic Scene Graph Dataset to Evaluate Rare Predicate Classes

Download the Haystack v1 archive here

The archive contains the following files/folders:

• annotation_v1.json: PSG-style dataset annotation. This means you could use the data as a drop-in replacement for PSG. However, this is not how the Haystack dataset was intended. Instead, you should definetely make use of our negative annotations.
• list_img.txt: List of RGB images from SA-1B
• Folders with this pattern: seg_sa_012345. The folder names correspond to the folder names from SA-1B

In addition to the standard PSG attributes, our annotation file also contains neg_relations (following the same format as relations). This is what makes Haystack special and is explained in our paper in more detail.

If you don't have the SA-1B dataset available, you can use this subset which only includes the required image files. Make sure that you have read and agree to the terms of the SA-1B Dataset Research License

For further questions, feel free to file an issue on GitHub or send an e-mail to julian.lorenz@uni-a.de