Visual Geometry Group - University of Oxford

James Philbin, Relja Arandjelović and Andrew Zisserman



Overview

The Oxford Buildings Dataset consists of 5062 images collected from Flickr by searching for particular Oxford landmarks. The collection has been manually annotated to generate a comprehensive ground truth for 11 different landmarks, each represented by 5 possible queries. This gives a set of 55 queries over which an object retrieval system can be evaluated.

Groundtruth Queries

The following image shows all 55 queries used to evaluate performance over the ground truth.

Image Montage

Flickr queries used to download images:

  1. All Souls Oxford
  2. Balliol Oxford
  3. Christ Church Oxford
  4. Hertford Oxford
  5. Jesus Oxford
  6. Keble Oxford
  7. Magdalen Oxford
  8. New Oxford
  9. Oriel Oxford
  10. Trinity Oxford
  11. Radcliffe Camera Oxford
  12. Cornmarket Oxford
  13. Bodleian Oxford
  14. Pitt Rivers Oxford
  15. Ashmolean Oxford
  16. Worcester Oxford
  17. Oxford

For each image and landmark in our dataset, one of four possible labels was generated:

  1. Good - A nice, clear picture of the object/building.
  2. OK - More than 25% of the object is clearly visible.
  3. Bad - The object is not present.
  4. Junk - Less than 25% of the object is visible, or there are very high levels of occlusion or distortion.

Terms and Conditions

The Oxford Buildings dataset consists of images provided by Flickr and use of these images must respect Flickr's Terms and Conditions of Use. In addition, use of the Oxford Buildings dataset must follow our Terms of Access. For privacy issues with the dataset, please refer to our Dataset Privacy Notice.

Downloads

All data needed for evaluation is given below:

  1. Groundtruth files
  2. C++ code to compute the ground truth

Additionally, we've made extra data available for the 5K dataset:

  1. README for the following files
  2. Compressed binary file of SIFT descriptors for the 5K dataset [MD5: fbc0e85c5065f6d97d519a7f2ed3e3f9]
  3. Compressed text files containing word IDs and geometry for the 5K dataset using a vocabulary size of 1M [MD5: 0734e6023f6bf2140b9531af22ce7953]
  4. Face annotations

Computing the average precision

  1. Compile the compute_ap.cpp file, using (on linux) g++ -O compute_ap.cpp -o compute_ap.
  2. To compute the average precision for a ranked list, rank_list.txt, one runs ./compute_ap christ_church_1 rank_list.txt, for the first Christ Church query, etc.

Relevant Publication


J. Philbin, O. Chum, M. Isard, J. Sivic and A. Zisserman
Object retrieval with large vocabularies and fast spatial matching  
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2007)

Acknowledgements

This work is supported by an EPSRC Platform grant.