Categorization and perceptual regrouping: Methodology and segmentation
Postdoctoral project
Vénérée Randrianarisoa
Robert Bergevin (Supervisor)
Problem: Research in cognitive psychology has established that even in the absence of high level knowledge of the observed scene, the human visual system naturally organizes the elements according to the laws referred to as the laws of perceptual grouping. Thus, the human visual system tends to group together two elements which are symmetrical, similar, close, parallel, etc… David Lowe derived his theory of non-accidentalness from this concept. This theory stipulates that primitives or groups of primitives having perceptual characteristics such as similarity, parallelism, symmetry, etc… are not presented randomly in the image but that they are derived from the same object or from a group of objects in the scene. The goal of perceptual grouping is thus to exploit the perceptual characteristics so as to extract the structural organization which represents an object or a group of perceptually significant objects. Most existing perceptual grouping methods use unitary or binary perceptual characteristics. Moreover, they follow an ascending hierarchy: thus the most prominent primitives respecting the laws of perceptual grouping are grouped to form a group which is supposed to be perceptually significant. The problem which arises in these approaches is that they do not propose global criteria which allow the evaluation of the global quality of the perceptually prominent group, since they essentially use local criteria related to the local perceptual characteristics. Moreover, very few studies using global criteria propose methods which treat large groups, i.e. formed by a large number of primitives. Most studies deal with simple groups such as convex groups, rectangles, simple polygons…
Motivation: Our approach differs from these approaches in that we propose to determine a set of global perceptual criteria which will allow a category of objects to be extracted. These criteria will enable the evaluation of the quality of the segmentation in order to determine if the group formed is perceptually significant and if it has a good perceptual shape.
Approach: Our approach consists in: (i) determining optimum criteria of an abstract object category by subjective human observations, (ii) forming global quality criteria defined by a human operator, (iii) extracting the primitives which optimize these global quality criteria. To determine these global criteria, we will use a methodology referred to as SAFE (Subjectivity And Formalism Explicitly). The goal of this methodology is to validate the judgement derived from formal calculations which we refer to as FGT (Formal Ground Truth) by human judgement which we call SGT (Subjective Ground Truth). Thus, the comparison of subjective human judgement with formal judgement will enable the use of a subjective ground truth which is useful for the evaluation and validation of the proposed criteria and algorithm. In order to implement the SAFE methodology within this project, we have developed a new tool, SAFE-T, with a Master’s student, Jean-François Bernier. This interactive tool will allow groups consisting of constant curvature primitives (circular arcs and straight line segments) to be manually and automatically generated.
Challenges: The detection and categorization of objects should involve the same criteria as the visual human system to classify an object as belonging to a given category. The extraction of a perceptually significant group (gestalt) is an important challenge but the perceptual evaluation of this group is even more challenging as it entails the evaluation of its proper shape represented in cognitive psychology by the “prägnanz”.
Applications: We will apply our approach in researching global perceptual criteria to find the silhouette of a category of objects, in this case multi-part objects. However, the approach could then be further refined to deal with the categorization of more specific objects. One of these applications is content-based image retrieval (CBIR).
Expected results: The results expected from this project are the definition and validation of a set of global perceptual criteria for the detection and categorization of multi-part objects.
Calendar: April 2004 - April 2005
Last modification: Oct 1 2007 1:47PM by veneree


©2002-. Computer Vision and Systems Laboratory. All rights reserved