Multiple-View Model Alignment MIT Artificial Intelligence Lab Overview We consider the problem of aligning a three-dimensional model of an object (left) to one or more images of that object (right). This problem arises in medical image registration and many computer vision applications. Single-View Alignment Existing methods use Mutual Information to align the model to a single view. While this method correctly aligns the model in plane, there is generally a gross translation offset along the ray of the camera. This drawback limits the feasibility of this method in precision applications. Registration We extended existing single-view registration methods to make optimal use of multiple-views simultaneously. The additional constraint locks down the position of the model in all three dimensions. The figures on the left and right show the 3D model with edges from the two images overlaid. Registration While single-view registration techniques can result in errors up to 25 mm, the error after performing a two-view registration was under 3 mm. Using the second view, the registration algorithm converges more quickly and from further initial positions. The two images (left and right) show points from the model in red overlaid.