NORB

NORB, short for New Object Recognition Benchmark, is a publicly available dataset created to evaluate object recognition algorithms under variations in lighting, viewpoint, and elevation. It was introduced in the early 2000s as a benchmark for 3D object recognition using grayscale stereo images.

The dataset comprises five object categories: airplanes, cars, motorcycles, ships, and trucks. Across these categories are 50 distinct objects (10 per category). Each object was photographed under several lighting conditions, elevations, and azimuths, yielding stereo image pairs for each view.

Images are grayscale and typically captured at 96×96 pixels. Each sample consists of a stereo pair with associated metadata, and the data are provided with predefined training and testing splits and labels for object identity, category, and pose.

NORB has been used to study 3D object recognition, view-invariant representations, and stereo vision in machine learning. It has served as a benchmark for evaluating early neural networks and, more recently, for testing deep learning models on smaller inputs with viewpoint and lighting variations.

The dataset is publicly available from official sources and is accompanied by baseline results and evaluation scripts. NORB remains a reference point in discussions of dataset design for object recognition under controlled but varied 3D viewing conditions, even as larger-scale datasets and more complex 3D benchmarks have emerged.