The Paris500k dataset consists of 501,356 geotagged images collected from Flickr and Panoramio. The dataset was collected from a geographic bounding box rather than using keyword queries. Thus, the images have a "natural" distribution, as shown in the figure on the right. The dataset is very challenging due to the presence of duplicates and near-duplicates, as well as a large fraction of unrelated images, such as photos of parties, pets, etc.
Recognition Ground Truth
We provide a ground truth for object recognition. Given a set of 13k object clusters discovered using Iconoid Shift, the task is to recognize the correct object in each of the provided 3k query images. The ground truth has exhaustive annotations for all query-object pairs. See our paper for further details. The dataset can be downloaded below.
Clustering Ground Truth
Additionally, we provide a clustering ground truth of 79 touristic landmarks covering 94,303 images. Please refer to this paper for details on its creation. Matlab scripts for evaluating cluster recall and cluster precision are provided as well. The ground truth is provided in the dataset package that can be downloaded below.
Notre Dame (10,292 images)
|Eiffel Tower (22,799 images)|
|Arc de Triomphe (8,115 images)||Palais de Chaillot (1,572 images)|
|Pantheon (1,149 images)||Musee d'Orsay (2,138 images)|
|Arc de Triomphe du Carousel (1,668 images)||Inside Notre Dame (1,109 images)|
|Invalides (1,137 images)||
Parc du Champs de Mars (1,100 images)
Images from Flickr come with a variety of licenses that should be respected when re-publishing the photos, e.g. in a publication. Copyright information can be retrieved using the photo ID using the Flickr API. Images from Panoramio may not be redistributed according to the Panoramio Terms of Service.
Dataset and Clustering Ground Truth
Due to copyright and space restrictions, we provide the dataset in the form of two download scripts written in Python which retrieve the images from Flickr and Panoramio automatically. Since some users might have deleted their photos from the sites, we also provide the full dataset on request. For this, please contact Tobias Weyand.
mmp_paris500k.tgz (16 MB)
We also provide the matching graph created by performing a pairwise tfidf matching and verifying the matches by fitting homographies using RANSAC. The matching graph is provided in the form of an adjacency list including tfidf scores and homographies. Using the provided Matlab script the adjacency list can be read into an adjacency matrix.
mmp_paris500k_matching_graph.tgz (620 MB)
Recognition Ground Truth
The purpose of this ground truth is to evaluate the precision of landmark recognition algorithms based on the output of a landmark clustering algorithm. See this paper for details.
paris500k_landmark_recognition_ground_truth.tgz (not available - 461 MB)