Region-Based Feature Interpretation for Recognizing 3D Models in 2D Images.

摘要

In model-based vision, features found in a two-dimensional image are matched to three-dimensional model features such that, from some view, the model features appear very much like the image features. The goal is to find the feature matches and rigid model transformations (or poses) that produce sufficiently good alignment. Because of variations in the image due to illumination, viewpoint, and neighboring objects, it is virtually impossible to judge individual feature matches independently. Their information must be combined in order to form a rich enough hypothesis to test. However, there are a huge number of possible ways to match sets of model features to sets of image features. All subsets of the image features must be formed, and matched to every possible subset of the model features. Then, within each subset match, all permutations of matches must be considered. Many strategies have been explored to reduce the search and more efficiently find a set of matches that satisfy the constraints imposed by the model's shape. But, in addition to these constraints, there are important match-independent constraints derived from general information about the world, the imaging process, and the library of models as a whole. These constraints are less strict than match-dependent shape constraints, but they can be efficiently applied without the combinatorics of matching. In this thesis, I present two specific modules that demonstrate the utility of match-independent constraints.... Model-based recognition, Image understanding, Vision, Object recognition, Region growing, Grouping, Recognition.