===================================================== | README file for | | Cornell 3D Location Recognition Datasets | | http://www.cs.cornell.edu/projects/3Dlocation/ | | v1.0, April 2012 | ===================================================== This document describes the contents and format of the Cornell 3D Location Recognition Datasets. This dataset is designed especially for algorithms for 2D-to-3D matching, and as such includes a 3D SfM model for each dataset. If you use this dataset in a publication, please cite the paper: Yunpeng Li, Noah Snavely, and Daniel P. Huttenlocher. "Location Recognition using Prioritized Feature Matching." Proceedings of ECCV 2010. Please read the following carefully before using the dataset, as there are several important notes. *See especially the special notes on the Rome16K dataset at the end of this document.* -== I. Contents of each dataset ==- Each dataset consists of a database -- including SIFT keys a 3D structure-from-motion (SfM) model -- and a set of query images, consisting of a set of SIFT keys. Out of respect for Flickr users, we are not distributing the original images with this archive, but we can send the images separately if you require access to them. (Many of the images are available on Flickr as well.) In particular, the provided files are as follows: list.db.txt -- the list of database image files list.query.txt -- the list of query image files db/ -- directory containing *squeezed* database SIFT key files (see below) query/ -- directory containing query SIFT key files bundle/ -- directory with 3D SfM models bundle/list.orig.txt -- full list of images used to build the SfM model bundle/bundle.orig.out -- original SfM model, before removing test images bundle/bundle.db.out -- database SfM model, after removing test images (use this model for testing if you would like to use our SfM results) bundle/components/ -- directory containing individual SfM models for different connected components of Rome (see Section IV below). -== II. Description of file formats ==- This section describes file formats for each of the data types above, with special notes as appropriate. A. List files (list.db.txt, list.query.txt). List files specify filenames to images in jpg format, one per line (keep in mind that the actual jpg files are not distributed unless requested). In addition, if the focal length of the image has been estimated from Exif tags, then that is also included. Images without known focal length information are specified with a line with a single field, the image name. Example: query/10970812@N05_2553027508.jpg Images with known focal length information are specified with a line with three fields: the image name, a zero, and the Exif focal length. (The second field is always zero but may change in future datasets.) Example: query/11289373@N03_2733280477.jpg 0 1280.00000 B. Key files We are providing (gzipped) SIFT key files for the database images and query images. One important note is that the database keys are *not* the original set of SIFT keys, but the set of SIFT keys after removing features that did not end up getting used for reconstruction (i.e., SIFT features that do not correspond to a point in the 3D model). We will make the *original* database key files available as a separate download (see the project webpage). The format of the key files is the same as that produced by David Lowe's SIFT extractor, described here: http://www.cs.ubc.ca/~lowe/keypoints/ C. 3D SfM models We provide two 3D SfM models, computed using Bundler. The formats of these files are described here: http://phototour.cs.washington.edu/bundler/bundler-v0.4-manual.html#S6 Description of files: - bundle.orig.out: Original bundle file, containing camera parameters for all database and query images. This can be used to compare recovered poses for query images to a "gold standard". - bundle.db.out: Database bundle file with query images removed (as well as any links between points and query images). This model is used during testing, to localize new images. Because it is registered with bundle.orig.out, camera parameters for images localized to bundle.out can be directly compared with those in bundle.orig.out. However, such comparisons should be take with a grain of salt, as the scale factor for the bundle files may not be meaningful. -== III. Accessing the original photos ==- While we are not distributing the original images with this archive, there are two ways to access the originals, if you need them. If you only need a few images (e.g., to show in a publication), then you can access photos directly on Flickr as follows: each photo filename has the form _.jpg. You can access Flickr photo through the Web with a URL of the form: http://www.flickr.com/photo_zoom.gne?id= e.g.: http://www.flickr.com/photo_zoom.gne?id=2000000000 If you need all of the photos (for instance, because you'd like to re-extract features), then please send a request to the email addresses below. -== IV. Special notes for the Rome16K dataset ==- This bundle files bundle.db.out and bundle.orig.out consist of several individual connected components of images from the city of Rome (e.g., the Trevi Fountain, the Colosseum, the Pantheon, etc.), merged into a single SfM model. However, these models are *not* individually aligned to a meaningful coordinate system; the positions, scales, and orientations of each individual model is more or less arbitrary. Hence, the geometry of the 3D model is not really that of Rome, although each individual model is self consistent. The SfM model can therefore be used to evaluate image registration rate, but cannot be used to evaluate metric accuracy of localization (unlike with the Dubrovnik16K dataset). Original reconstructions for the individual connected components are also included as separate list and bundle files in bundle/components/. Questions? Email snavely@cs.cornell.edu and yuli@cs.cornell.edu.