MPEG data set/base for Compact Descriptors for Visual Search (CDVS)

Author: MPEG

Partner: No

Contact: For MPEG CDVS project: Giovanni Cordara (




Total: 34000


List of databases currently considered for CDVS consists of 7 datasets, the two famous ZuBuD and the university of Kentucky DB, and 5 produced by the proponents. ZuBud database contains pictures of 200 buildings in Zurich, shot by 2 cameras under different viewing conditions (1000 reference images, 115 query images). University of Kentucky DB contains 10200 images, 4 views of object, 2550 objects, first image of each object is a query image. Stanford DB has 1200 objects (8 categories), 3300 query images. ETRI DB has 39 objects, captured with 22.5 degree changes in viewpoints, captured by using 4 cameras, and having a total of 3800 images 14 buildings, 150 query images. Peking University DB has 198 buildings, 13179 images (query or reference). Telecom Italia DB has 180 landmarks/buildings, 1440 images, 180 videos of each building, of 10 seconds each. Telecom SudParis DB has 35 buildings, 600 images (query or reference). In total the databases cover over 34K of images.


ZuBud Database TAR-GZIP (~486MB) ZuBuD Query Images TAR-GZIP (~3.1MB) University of Kentucky Recognition Benchmark Images ZIP (~2GB) Stanford Database