ZuBuD+

ZuBuD is an image dataset, used in object recognition problems.

  • It composed by 1005 images of 640×480 pixels, subdivided in 201 classes
  • The query images are only 115 of 320×240 pixels.

Dataset description

ZuBuD+, created in February 2017, introduces many query images balancing the class evaluated from the previous dataset.

Now the test set is composed by 1005 images of 320×240 pixels and all the classes are represented in the ground truth 5 times. The new images are obtained from the training images, choosing randomly an image from the five of every class, and then they are transformed using resize or rotation (±90°) and resize.

Sample image below

Download

The new dataset is available for the download (545Mb).

The folder includes:

  • the database images
  • the new query set balanced
  • the ground truth file updated
  • a python script that allows you to check the performance of your system (with an example).

State-of-the-art on ZuBuD+

  • intra-dataset setup (K=4281, D=128): 4,543
  • inter-dataset setup (K=128, D=64): 4,447 (the vocabulary is created using the features obtained by the Paris dataset).

The metric is used during the experiments is 5 x Recall @ top5.

What is “5 x Recall@top5”?

It’s the mean of how many images of the 5 retrievals are in the top 5 of the relative query.

Copyright

If you used this dataset please cite the following paper: A location-aware embedding technique for accurate landmark recognition.

Bibtex

@article{magliani2017location,
  title={A location-aware embedding technique for accurate landmark recognition},
  author={Magliani, Federico and Bidgoli, Navid Mahmoudian and Prati, Andrea},
  journal={arXiv preprint arXiv:1704.05754},
  year={2017}
}

For any question please contact Federico Magliani.