Montreal Scale-Change Dataset

The Montreal Scale-Change Dataset consists of sets of images of twelve different urban scenes within and around the downtown campus of McGill University in Montreal, QC. For each scene, images were taken at two or three distances from the scene's central object, with minimal changes in viewing angle, to capture the scene over a wide range of visual scales. These images form a total of 31 image pairs, which exhibit changes in foreground scale ranging from a factor of 1.5 to a factor of 6, with one exception in which the scale change factor is approximately 14. All images were taken with the rear camera of a Samsung Galaxy SIII phone. Ground truth is provided for each image pair in the form of a set of ten manually-annotated point correspondences distributed approximately evenly across the nearer image in the pair.

Data and Code

This dataset was used in the preparation of our papers Long-Distance Loop Closure Using General Object Landmarks and Scale-invariant Localization Using Quasi-semantic Object Landmarks to measure the performance of our proposed visual localization technique over significant scale changes. The IROS paper can be found on arXiv. The code used to generate our results for this dataset can be downloaded here.

The complete dataset can be downloaded here. If you make use of this dataset or the related code, please cite us:

@article{holliday2021scale,
  title={Scale-invariant localization using quasi-semantic object landmarks},
  author={Holliday, Andrew and Dudek, Gregory},
  journal={Autonomous Robots},
  volume={45},
  number={3},
  pages={407--420},
  year={2021},
  publisher={Springer},
}

Sample Scene Sets