Experimental Evaluation of Projective Invariants in Computer Vision
YI Yinghu · 哈尔滨工业大学学报：英文版 · 1998
In model-based computer vision, the fundamental difficulty in recognizing objects from images is that the appearance of a shape depends on the viewpoint. This problem is entirely avoided if the geometric description of the object is unaffected by the imaging transformations. This description is an invdriant. Four commonly used planar projective invariants encountered in practical applications are discussed and their reliability is illustrated experimentally.They are analyzed from a theoretical viewpoint to derive expressions which characterize the propagation of error and from an experimental viempoint by constructing some test patterns with known values of the invariants and comparing these values with those calculated from the image data. An algorithm using invariant to recognize objects under projection is also introduced. It is found that the accuracy and stability of the invariants are directly dependent on the performance of the feature extraction scheme employed to recover the test pattern geometry from the images.