Dense matching for multi-scale images by propagation

This article puts forward a new algorithm of dense matching between two images of the same scene, not necessarily stereoscopic and which can be of different scales. This algorithm is based on points of interest identified by a multi-scale detector and later matched according to a very high threshold to keep only the most reliable points. Afterwards, these points serve as germs for the next iteration. This propagation process is guided according to geometric constraints, and repeated until all the possible correspondents between the two images are obtained. The results of the experiments obtained on test images and those of the real world are very satisfactory even in difficult cases of great displacements and changes in appearance between the two captures.