SunHo Lee
Electrical Engineering
Ph. D. Candidate
(Visiting Researcher)

3D Object Representation and Recognition Using Superquadric Models From Range Images

Research objectives:
Three-dimensional (3D) object recognition is one of very important tasks which needs to be performed in many industrial applications of machine vision and research areas, such as robotics and computer vision[3,4]. Obviously, complete model description of 3D object and elaborate matching algorithm are critical steps in building an efficient object recognition system. In the past, many 3D object representation schemes have been proposed for computer vision systems, for example wire-frame representation, generalized

   
Fig.1 CAD model of input object (VRML File)
You must have Cosmo to view the VRML images.


Fig.   Curvature-based range image segmentation (column object.avi)

   

cylinder representation, multi-view representation and other volumetric representations. Among them, volumetric representation is relatively advantageous because it is viewpoint independent. In general, human visual system perceives 3D objects intuitively by decomposing into the object parts based on the theory of Recognition By Components (RBC) In this point of view, one of most appropriate representations is superquadric model. This superquadrics has the advantage that it has the capability to describe a wide variety of primitive shapes only with a finite number of parameters including translation, rotation, and global deformation.

This research presents an integrated recognition scheme by using both surface and volumetric information. In this algorithm, segmentation provides surface information of the scene object, and we can achieve part decomposition. In the next step, we extract superquadric parameters and the matching features including the number of volumetric primitives and the junction relationships between them. In addition, this method calculates junction vectors between neighboring primitives, which indicate a distance and direction between the centroids of primitives. In the final step, we match the feature values of the scene object with those of the model objects. Fig.1 is CAD model of input object. Fig. 2 and Fig. 3 show curvature-based surface segmentation and part-based superquadric model shape recovery.

This research presents an integrated recognition scheme by using both surface and volumetric information.  In this algorithm, segmentation provides surface information of the scene object, and we can achieve part decomposition.  In the next step, we extract superquadric parameters and the matching features including the number of volumetric primitives and the junction relationships between them.  In addition, this method calculates junction vectors between neighboring primitives, which indicate a distance and direction between the centroids of primitives.  In the final step, we match the feature values of the scene object with those of the model objects.  Fig. 1 and Fig. 2 show curvature-based surface segmentation and part-based superquadric model shape recovery.

Methodology and Results:
In general, three major steps are required to establish a 3D model-based object recognition system using superquadrics as object models: database building, scene object model recovery and model matching. First, we need to construct a superquadric database for efficient access of a large number of object models. Then. the model of the target object is accurately recovered from the range sensor data. Most of conventional researches using superquadrics are mostly concentrated on 3D shape recovery and segmentation of range images through superquadric parameter estimation. After these two steps have been completed, the last step of the system is to correctly and efficiently match a recovered superquadric with a set of superquadrics in the model database.

In spite of many advantages, we confront difficulties in recognizing 3D object by using superquadrics alone. This paper summarizes these reasons as following. At first, superquadrics cannot provide a uniform representation of 3D object, then it is too difficult to build model base constantly. Secondly, volumetric representation is a global model and is insensitive to local variations which has no surface information: surface type and relationship. The use of additional surface information is advantageous for a precise recognition.

This research shows that part-based superquadric model (PBSM) can be used as an effective approach to 3D recognition. Since superquadrics is one of volumetric model representations, our method is very similar to human perception, so robust to non-linear shape changes according to viewpoint. Furthermore, the use of both junction vectors between neighborings and surface adjacency graph (SAG) can solve self-occlusion problem.

This work is being conducted at IRIS lab by SunHo Lee under the supervision of M. A. Abidi (Thesis Chair). This work is supported by DOE's University Research Program in Robotics under grant DOE-DE-FG02-86NE37968.