Computer Vision News - December 2016

The main technique to do that is features detection: bifurcations of blood vessels are taken as landmarks and guide us to stitch the images, just like the ophthalmologist would do if there was no montage software. Even after that, not all matchings will be accurate yet: techniques like RANSAC are used to add consistency and eliminate incorrect matchings. Points of interest, even though correctly detected in every image, cannot be correctly paired yet and it is interesting to understand the reason for that. Generally, when we want to match and stitch, it is important to understand what transformation we expect: naïve stitching will generate rigid transformation and even some minor rotation. But since the eye is a 3D sphere, its shape requires a more elaborated transformation in order to generate a correct stitching of the retina images in the space: in other words, each image is a projection of parts of a sphere over a 2D detector in the camera. We do the inverse transformation, from 2D back to 3D images , which are ready to be stitched only when this process is done correctly. Inaccurate transformation would result in duplication and breaks of blood vessels, for a very inaccurate panoramic view. The task of RANSAC is to make sure that this transformation is uniform and that all the paired matchings are true. Once this process is completed with the required attention, the exact montage of retinal images will be ready to be presented to the ophthalmologist in a 2D format. Project Computer Vision News Project 27