Access this dataset

Our read-only mount feature is the fastest way to access repositories of any size, streaming files to your machine in seconds. Mount provides a file-system view of repositories for smooth access from any application, whether you're using notebooks, code, or your Finder window.

If you haven't already, install and authenticate, then dive in to see what's in this 4GB Flickr dataset.

Mount

If you installed with PyXet, run:

xet mount xet://XetHub/Flickr30k/main Flickr30k

If you installed with Git-Xet, run:

git xet mount https://xethub.com/XetHub/Flickr30k.git 

You can copy these commands for any repository by clicking the purple Access button right above this README.

Explore

Our Flickr30k dataset includes a 13.9MB results.csv file that lists 5 annotations per image, as well as around 32k images organized by the first two numbers of each file name. Phew.

Dust off your favorite file or photo browser to navigate to 13/131090759.jpg, which shows how we all feel when we have to lug our big files around to do our jobs. And while you're at it, click around some more to see the randomness of public internet images.

Next steps

🐛 This dataset is not perfect. Check out the pull requests to see what collaborative dataset review looks like.

🎓 Read more about streaming access.

🛠️ Move on to the next step of the Quick Start to make your first changes in XetHub!

About this dataset

The Flickr30k dataset has become a standard benchmark for sentence-based image description. This paper presents Flickr30k Entities, which augments the 158k captions from Flickr30k with 244k coreference chains, linking mentions of the same entities across different captions for the same image, and associating them with 276k manually annotated bounding boxes. Such annotations are essential for continued progress in automatic image description and grounded language understanding. They enable us to define a new benchmark for localization of textual entity mentions in an image. We present a strong baseline for this task that combines an image-text embedding, detectors for common objects, a color classifier, and a bias towards selecting larger objects. While our baseline rivals in accuracy more complex state-of-the-art models, we show that its gains cannot be easily parlayed into improvements on such tasks as image-sentence retrieval, thus underlining the limitations of current methods and the need for further research.

License: CC0 1.0