NORB Dataset is an experimental dataset for shape-based 3D object relocalization, which contains 50 toy images of 5 general categories: quadrupeds, people, airplanes, trucks, and cars. The objects in the dataset are imaged by two cameras under 6 lighting conditions, 9 altitudes, and 18 azimuths. It consists of a training set containing 5 instances and a test set containing 5 instances.

This dataset was released by New York University in 2004. The main publishers are Fu Jie Huang and Yann LeCun. The related paper is "Learning Methods for Generic Object Recognition with Invariance to Pose and Lighting".