Multispectral interest points for RGB-NIR image registration
This project aims to extend standard feature detectors to use an additional near-infrared (NIR) channel to improve matching performance.
Light reflected from an scene can be divided into visible light and near-infrared (NIR) light. The visible part is what we see, with a wavelength range of 400-700nm, and NIR is invisible to the naked eye, with a wavelength of 700-1100nm. For more information, please read Clement Fredembach's page on near-infrared.
We used this additional channel to extend standard feature detectors: Harris corner detector and Difference of Gaussians (DoG). Their extension consists of combining the RGB and NIR channels in a way that brings out strong features from every channels.
Resources
The source code for running the tests will be soon available here, as well as the dataset.
