Our interest points are located at extrema of the Laplacian of the image in scale-space. This function is chosen for its response at points with 2-dimensional structure, and for the fact that it can be implemented very efficiently using a Laplacian Pyramid [4]. In a Laplacian Pyramid, a difference of Gaussians is used to ap- proximate the Laplacian. Pyramid representations have the advantage that the minimum number of samples are used to represent the image at each scale, which greatly speeds up computation in comparison with a fixed resolution scheme. To find the maxima and minima of the scale-space Laplacian we first select samples which are extrema of their neighbours ±1 sample spacing in each dimension. Then, we locate the extrema to sub-pixel / sub-scale accuracy by fitting a 3D quadratic to the scale-space Laplacian