Spectral intersection over union: a bounding box overlap metric for hyperspectral object detection

Abstract

Hyperspectral images provide significant spatial and spectral information which are widely used in object detection. Two-stage detectors are commonly employed in hyperspectral object detection, where effective region proposals play a crucial role in accurate object localization. However, during non-maximum suppression (NMS) process, the Intersection over Union (IoU) metric based solely on spatial geometric information is inadequate for discriminating between similar proposals. This results in a substantial number of expected proposals with dissimilar characteristics are eliminated. In this paper, we analyze the spectral information in hyperspectral images to distinguish the characteristics of different proposals. Furthermore, this paper proposes the Spectral IoU (SIoU) by introducing spectral signature differences as a new metric. This improves the ability to differentiate between different object instances and increases the recall rate of bounding boxes with high localization confidence in region proposal stage. Moreover, SIoU can be simply integrated into the hyperspectral objection detection frameworks without introducing additional computational complexity. Extensive experiments on the Semi-Supervised Hyperspectral Object Detection Challenge dataset demonstrate the effectiveness of our method.