Open Dataset and Enhancement Method for Long-wave Thermal Diurnal Material Classification

Thermal infrared imaging is widely used for detection when lighting conditions are expected to be inconsistent or low intensity. Because infrared sensors capture different wavelengths, they are unaffected by changes in lighting conditions. Despite this, an object that appears hotter at one time of day might not appear hotter at another time due to different heating and cooling rates. Until now, only two diurnal thermal datasets are openly available. Our datasets differ in that ours is on a different scale (short-distance targets vs. aerial land surveys) and focuses on different materials that would likely appear in artificial structures (glass, metal, plastic, and styrofoam) instead of on different types of plants to reflect performance in an urban environment better. Our dataset consists of over 72,000 thermal images of a static scene taken over about one and a half months. We expect our dataset to provide researchers with data to train networks to detect objects accurately during crossover events. In addition, we provide an image preprocessing method to provide the network with classical computer vision information to improve the classification of materials over the diurnal cycle.