Maarten W. Borggreven and Shinjiro KAWATO
Implementation of Feature Tracking and Factorization Algorithm for Shape and Motion Recovery from Image Streams
Abstract:In 1990 Tomasi & Kanada started a series on Shape and Motion Recovery from Image Streams under orthographic projections. This report contains a first implementation and analysis of the main algorithms in order to provide a solid basis for own research on this topic. First aspect is the selection of feature points, or rather feature windows. Selection is done on basis of thresholding eigenvalues of the coefficient matrix, containing the gradient vectors of the feature window. Second aspect is the tracking of the feature windows along the image stream. Final aspect is the actual 3D Shape recovery using a proposed factorization method. Singular Value Decomposition is used to decompose the measurement matrix into shape and motion. The method further calls upon the rank principle for noisy images to remove redundancy from the measurements. A few experiments are done to analyse the performance and create thoughts about future research.