A Comprehensive 2D + 3D Dataset for Benchmarking Hyperspectral Imaging Systems

Abstract

Hyperspectral images are represented by numerous narrow wavelength bands in the visible and near-infrared parts of the electromagnetic spectrum. As hyperspectral imagery gains traction for general computer vision tasks, there is an increased need for large and comprehensive datasets for use as training data. Recent advancements in sensor technology allow us to capture hyperspectral data cubes at higher spatial and temporal resolution. However, there are few publicly available multi-purpose hyperspectral datasets captured in outdoor terrestrial conditions. Furthermore, there are no publicly available datasets that include 3D mesh representations of objects captured in outdoor scenes. This article introduces the first hyperspectral dataset of 3D objects and terrestrial outdoor scenes, the Tufts Outdoor Hyper-spectral Dataset (TOHS Dataset). The dataset includes 100 2D + 3D hyperspectral scenes, each containing 164 spectral bands. The contributions of this work are 1) Detailed description of the content, acquisition procedure, and benchmark results on state-of-the-art neural networks for 3D object scenes in the Tufts Hyperspectral Database; 2) The first-of-its-kind hyperspectral 3D dataset of outdoor objects that will be publicly available to researchers worldwide, which will allow for the assessment and creation of more robust, consistent, and adaptable AI algorithms; and 3) a comprehensive and up-to-date review on hyperspectral systems and datasets.