Geometric Methods for ATR: Shape Spaces, Metrics, Object/Image Relations, and Shapelets.

摘要

This report details our recent progress (1 Sept. 2006 to 30 Sept. 2007) and summarizes the work done over the entire 40 months of the project. It outlines our recent success in creating the first 'global object/image equations and metrics for the full perspective case, and it discusses some of the additional experimental testing we did to verify the robustness of our algorithms. We extended our analysis of the relationship between the shape spaces for point features under similarity transformations and those under the affine group. The former are relevant to orthographic sensors (radar) and the later arise when dealing with weak perspective sensors (optical - far field). Understanding the relationship between the two types of shape spaces facilitates fusing data from these two types of sensors. In addition, we worked on global embeddings of the shape spaces in the orthographic (radar) case, and we continued to work on metrics for the case where the features are taken as unordered collections of points. Currently, we are working on the 3D reconstruction problem, especially shape from motion and have started to look at advanced statistical/noise issues.