Stereo-image rectification for dense 3D reconstruction in scanning electron microscope

Stereo rectification is a crucial step for a number of computer vision problems and in particular for dense 3D reconstruction which is a very powerful characterization tool for microscopic objects. Rectification simplifies and speeds up the correspondence search in a pair of images: the search space is reduced to a horizontal line. It is mainly developed for perspective camera model: a projective transformation is found and applied to both images. This paper addresses the rectification problem for an image pair obtained with Scanning Electron Microscope (SEM). In this case, image formation is described by a parallel projection, indeed perspective effects can be neglected because of the low value of the sample size over working distance ratio. Based on these hypotheses, a robust estimation of the fundamental matrix, that describes the geometry of the image pair, is proposed. It filters out up to 50% of outliers in a feature correspondence set. Then, the matrix is used to develop a rectification solution: the problem is reduced to two rotations about the optical axis of the camera. The solution is accurate and does not require any calibration of SEM. It is validated with two image pairs from two different field effect SEMs.