SUN database: Large-scale scene recognition from abbey to zoo

This paper addresses scene categorization, a fundamental problem in computer vision whose study had been constrained by the narrow scope of available datasets—while object-recognition databases had hundreds of classes, the largest scene dataset contained only 15 categories. To remedy this, the authors introduce the extensive Scene UNderstanding (SUN) database, which contains 899 scene categories and 130,519 images.

Using 397 well-sampled categories, the authors evaluate numerous state-of-the-art scene-recognition algorithms and establish new bounds on performance. They also measure human scene-classification accuracy on the SUN database and compare it against the computational methods, and they study a finer-grained scene representation to detect scenes embedded inside larger scenes, providing both a large benchmark and a reference for human-versus-machine scene understanding.