Unsupervised calibration of RGB-NIR capture pairs utilizing dense multimodal image correspondences

In this paper, we propose an unsupervised calibration framework aimed at calibrating RGB plus Near-InfraRed (NIR) capture setups. We favour dense feature matching for the case of multimodal data and utilize the Scale-Invariant Feature Transform (SIFT) flow, previously developed for matching same-category image objects. We develop an optimization procedure that minimizes the global disparity field between the two multimodal images in order to adapt SIFT flow for our calibration needs. The proposed optimization substantially increases the number of inliers and yields more robust and unambiguous calibration results.