Learning multiscale image models of 2D object classes

Abstract

This paper is concerned with learning the canonical gray scale structure of the images of a class of objects. Structure is defined in terms of the geometry and layout of salient image regions that characterize the given views of the objects. The use of such structure based learning of object appearence is motivated by the relative stability of image structure over intensity values. A multiscale segmentation tree description is antomatically extracted for all sample images which are then matched to construct a single canonical representative which serves as the model 0fthe class. Different images are selected as prototypes, and each prototype tree is refined to best match the rest of the class. The model tree for the class is that tree which is best supported over all the initializations with different prototypes.Matching is formulated as a problem of finding the best mapping from regions of example images to those of the model tree, and implemented as a problem in incremental refinement of the model tree using a learning approach. Experiments are reported on a face image database. The results demonstrate that a reasonable model of facial geometry and topology is learnt which includes prominent facial features.

Region-Based Hierarchical Image Matching

“…For the most part, existing methods use greedy approaches, where, for example, matching is done top-down only between regions at the same tree level, such that a bad match between two regions penalizes attempts to match their respective descendants (Cohen et al 1989b;Perrin et al 1998). In Glantz et al (2004), a sequence of planar graphs, representing multiscale image segmentations, is defined as segmentation hierarchy, and ten topological relationships between any two regions v 1 and v 2 in the segmentation hierarchy are defined as a combination of the following basic relationships: (i) v 1 may be adjacent to v 2 , (ii) v 1 may enclose but not contain v 2 , (iii) v 1 may contain v 2 , and (iv) v 1 and v 2 may be apart.…”

Unsupervised Category Modeling, Recognition, and Segmentation in Images

“…Another approach to the automatic construction of object shape models recursively merges pairs of primitive curve elements that satisfy a set of user-specified generalization criteria [31]. In [32], a hierarchical category model is incrementally refined through matching the segmentation trees of a given set of images with the model, where matching is done top down in a greedy manner, only between regions at the same tree level, such that a bad match between two regions penalizes attempts to match their respective descendants. In [33], a tree model of an object shown in a given input image is learned by matching the input image to a sequence of templates provided by the user.…”

The Evolution of Object Categorization and the Challenge of Image Abstraction

“…However, in a more general recognition task, more ambitious domain-independent grouping is essential, which clearly introduces additional complexity. To help manage this complexity, feature hierarchies have re-emerged, in combination with powerful learning tools, to yield exciting new categorization frameworks [7,6,179,41,241,92,171,4,242,273]. 3 Figure 8 illustrates the system of Todorovic and Ahuja [242], in which a region-based hierarchical object model is learned from training examples and used to detect new instances of the model in query images.…”