Illumination-insensitive Binary Descriptor for Visual Measurement Based on Local Inter-patch Invariance

Abstract: Binary feature descriptors have been widely used in various visual measurement tasks, particularly those with limited computing resources and storage capacities. Existing binary descriptors may not perform well for long-term visual measurement tasks due to their sensitivity to illumination variations. It can be observed that when image illumination changes dramatically, the relative relationship among local patches mostly remains intact. Based on the observation, consequently, this study presents an illumination-insensitive binary (IIB) descriptor by leveraging the local inter-patch invariance exhibited in multiple spatial granularities to deal with unfavorable illumination variations. By taking advantage of integral images for local patch feature computation, a highly efficient IIB descriptor is achieved. It can encode scalable features in multiple spatial granularities, thus facilitating a computationally efficient hierarchical matching from coarse to fine. Moreover, the IIB descriptor can also apply to other types of image data, such as depth maps and semantic segmentation results, when available in some applications. Numerical experiments on both natural and synthetic datasets reveal that the proposed IIB descriptor outperforms state-of-the-art binary descriptors and some testing float descriptors. The proposed IIB descriptor has also been successfully employed in a demo system for long-term visual localization. The code of the IIB descriptor will be publicly available.