Generic Recoginition of the Main Object in a 2D Scene by Studying the Relationships Between the Contour Primitives
Alexandre Filiatrault
Robert Bergevin (Supervisor)
Problem: Generic recognition without a priori knowledge of the subjects in an image has been an area of interest for several years in computer vision. The difficulty lies in finding the specific characteristics of a real object which differentiates this object from its immediate surroundings. These characteristics often differ from one image to another, which leads us to use more than one type of discrimination at a time to improve the segmentation and better isolate the contour of the main object in a scene with respect to the interior of this object and the background of the scene.
Motivation: An ideal generic computer vision two dimensional recognition system includes three essential elements: the segmentation of the regions of interest, the categorization of these regions into entities according to global characteristics and the association of these entities with one of those in an already established bank of knowledge. Each of these three steps is essential to ensure that the system functions properly. The goal of this research project is to contribute to the improvement of the segmentation step.
Approach: This project uses the SAKE methodology which consists in:
  • Defining the expected results for each type of image.
  • Choosing the images on which the algorithm will be tested.
  • Choosing subjective criteria for evaluating solutions.
  • Determining formal mathematical criteria for evaluating solutions.
  • Designing, implementing and applying the resolution algorithms required to solve the problem on the chosen images.
  • Rating the solutions obtained and comparing these with formal ratings.
  • Modifying the resolution algorithm if necessary.

The approach proposed for identifying the important parts of an image rely mainly on the study of the contour primitives. The procedure consists in using the characteristics of these primitives so as to regroup them and then give them an order of importance. Regrouping data includes the colour, position, orientation and length, and for the primitive of an arc of a circle it also includes the radius.

The treatment studied herein proceeds in an iterative manner so as to converge to a solution where all of the primitives are well-balanced in their relationships with the others. This solution must make the best possible discrimination so as to exclude a maximum number of non-desirable primitives, representing textures or parts of the background, while keeping a large majority of data enabling the identification and characterization of the main object. This approach is preferred in view of being combined with others to obtain superior results with the help of complex input images. It will then be possible to characterize the objects found and identify them with the help of a bank of knowledge.

Challenges: The inherent challenges in such an approach are numerous and complex. Algorithms must be developed which efficiently isolate the main object of a scene while keeping in mind that the characteristics which are common to the primitives to be kept can differ largely from one image to another. One must also take into account the noise present in any computer image, the limited resolution of the images, the amount of memory available in the treatment tool and the speed of the computation tool. Another challenge involves the solution of problems related to the slow changes in intensity. These are less inclined in producing primitives and can thus hinder treatments.
Applications: The development of a mechanism enabling the search for a main object is an important step in generic recognition without prior knowledge. This ability to efficiently automatically recognize ones environment is very useful in areas such as robotics and virtual reality.
Calendar: September 2004 - December 2007
Last modification: Oct 1 2007 1:40PM by afiliatr


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