Off-campus UMass Amherst users: To download dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.
Non-UMass Amherst users, please click the view more button below to purchase a copy of this dissertation from Proquest.
(Some titles may also be available free of charge in our Open Access Dissertation Collection, so please check there first.)
Matching two-dimensional images to multiple three-dimensional objects using view description networks
Abstract
This dissertation explores the problem of effectively matching a single 2D image of a potentially cluttered scene to a library containing multiple polyhedral objects. The contributions of the research are in three areas: (1) the analysis of the usefulness of 2D features for 3D object discrimination under view variation; (2) the automatic compilation of a network of object view representations for the efficient matching of multiple 3D objects in 2D images; and (3) the design of a recursive indexing process for matching images to networks of view representations. In the analysis of 2D features for 3D object discrimination under view variation, invariants with respect to view position are studied. The recognition of 3D objects from a single, unknown view can be facilitated by the use of image features that are invariant. However, we have shown that there does not exist a function that is view-invariant for arbitrary 3D point sets of size n, for any n, given standard models of projection. Given this result, systems designed for unrestricted recognition problems cannot be expected to be as effective as those for specialized domains. In addition to this result, our analysis includes an empirical study of the variation, with respect to view, of some important 2D features defined for the projections of 3D line segments. The second area of contributions is in the compilation of object representations for their efficient matching in 2D images. The representation scheme emphasizes the use and organization of object 2D view descriptions into networks. This is in contrast to 3D model networks, and the organization of object information into tree structures. In addition, a method of automatically constructing and organizing view representations for multiple objects is developed that integrates view analysis and multi-object discrimination analysis. The third area of contributions is the development of a recursive indexing process for matching images to view description networks. With this approach, the evidence from matches to multiple parts of the image and the convergent structure of the description network are used together to provide a focused search for the correct interpretation. Experiments on real, digital images of cluttered scenes are presented that help to demonstrate this. In these experiments, there were tens of billions of potential matches between the images and the objects in the library to be recognized. In spite of this, the number of 3D matches actually hypothesized was very small, averaging 1.6 hypotheses per correct 3D match.
Subject Area
Computer science
Recommended Citation
Burns, John Brian, "Matching two-dimensional images to multiple three-dimensional objects using view description networks" (1992). Doctoral Dissertations Available from Proquest. AAI9219411.
https://scholarworks.umass.edu/dissertations/AAI9219411