2. Establish keypoint correspondences between successive images. As we showed in [17] the reprojection errors can be decreased if the matching is performed with two predecessors, instead of only one. Because of the high dimensionality of the keypoint descriptor, comparing two keypoints regarding their similarity is computationally expensive. To circumvent the quadratic complexity for comparing every keypoint in one image with every keypoint in another image, we use a bin-based approach: we divide the image space into bins and add each keypoint to the corresponding bin. The matching is then performed only with the other image’s keypoints in the bin at the same position and its neighbours.