This is the public release of the Scape-Imperial Localisation Dataset (SILDa).
Please Note. This is work in progress, and we are currently adding examples, and updating the documentation. If details for a specific SILDa task are not here yet, please check again soon.
We provide a bash script download.sh to download all the available data for SILDa.
To execute, simply open a terminal and type sh download.sh. Please note that the download
will take some time due to the amount of data (approx. 60GB).
For the local patches task, we provide a set of 557166 interest
points, each consisting 7 patches. This leads to a total of 3900162
individual patches. Note that the patches are saved in their full
color format, to enable experiments with colour descriptors.
Descriptors extracted from RGB patches have not been extensively
explored in the deep learning literature, possibly due to the fact
that no large scale colour patches dataset is widely available.
To submit your method's results to the local patches task, please check the silda-patches notebook where all the required steps are described in detail.
Results will be based on patch retrieval accuracy, using a method similar to the HPatches retrieval protocol.
For the image matching task, we provide a set of 335k pairs of images.
To submit your method's results to the image matching task, please check the silda-matching notebook where all the required steps are described in detail.
Results will be based on computing matching accuracy, using epipolar geometry.
For the camera pose estimation task, we provide a set of 6064
query images for which the camera pose is not known. We also provide
8344 images with known poses as a training set. The users can
utilise any method they want, to produce full 6DoF camera poses for
the unknown queries.
To submit your method's results to the camera pose estimation task, please check the silda-camera-poses notebook where all the required steps are described in detail.
Results will be based on camera pose accuracy, i.e. measuring translation and rotation errors between the prediction and the ground truth. For more information on this task, please see www.visuallocalization.net
For the building recognition task, we provide a set of 6064 query
images for which the observed buildings are not known. We also provide
8344 images together with the labels of the observed buildings. We
provide a total of 25 buildings, and we provide on average 300 images
per building as training set. This can be though as a few shot
learning task. The users can utilise any relevant method, to produce
building labels for the query images.
To submit your method's results to the camera pose estimation task, please check the silda-building-recognition notebook where all the required steps are described in detail.
Results will be based on standard multi-class classification mAP measurements, for the building recognition task.
More details for this task will be available soon.
More details for this task will be available soon.
The patches, matching and 6DoF tasks are parts of challenges associated with 2 CVPR 2019 workshops. For more information, please refer to the individual websites below.
CVPR 2019 Workshop on Image Matching
CVPR 2019 Workshop on Long-Term Visual Localization under Changing Conditions
The images of SILDa are licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License and are intended for non-commercial academic use only.
We are not currently planning to make the dataset available for commercial use.
By using SILDa, you agree to the license terms set out above.
We take privacy very seriously. For this reason, we used software to automatically blur faces and licence plates, and in addition we verified the results manually. If you have any concerns regarding the images and other data provided with this dataset, or find faces or licence plates that we have missed, please contact us.