Project goal and description
The goal of the project is to build a system for learning functional reasoning from range images. The recognition process is conducted via constructing a generic internal representation of the objects viewed in the image. The internal representation consists of a multi-level hierarchy of functional parts that is constructed by no a-priori knowledge during the learning stage. The functional parts, as well as their relationships, are recognized employing a large set of geometric properties. The problem of learning is solved by building a distribution probabilities framework for geometric attributes.
The proposed system consists of two phases: a learning phase and a classification one. In the first phase, several instances of object class are presented to the system. The objects are segmented and labeled into constituents: primitive and functional parts. The learning stage computes the values of the geometric attributes of the constituents and the relationships between them.
Once the first phase is finished, the system is considered to be able to classify the learned object class. Our final target is to give grades to the objects being classified. The grades reflect the level of fitting the classified object to ones that the system has learned in the first phase.
The recognition process employs an analysis of the detected primitive parts themselves and the relationships that exist among the detected primitive parts. We have generalized the mechanism of decomposition into parts and primitive-to-primitive connections to a multi level approach in the following sense. Each primitive part or group of primitive parts and primitive-to-primitive connections among them that can fulfill a certain functional task are classified as a functional part. Further, several functional parts and the relationships among them can define a functional task and can form a higher level functional part. This hierarchy can be as complex as the user allows. This approach is known in the literature as recognition by functional parts.
An example of the hierarchy could be an armchair which can be modeled as a chair with arms. The arms represent a simple functional part with the functionality of supporting the arms, while the chair is a high level functional part because it describes a more complex functionality. The chair is modeled as a stool with a back support whereas the stool consists of a sittable part, a ground support part, and a free space to sit in.
A relationship between a pair of functional parts is called a functional-to-functional connection. Connection can be of two types: primitive-to-primitive or functional-to-functional.
The following diagram describes connections between 3 functional parts. While the functional or primitive parts are represented by nodes the relationships between them are represented by edges. Each level in the functional hierarchy has a clique structure and each pair of functional parts are characterized by a relationship expressed in terms of geometric attributes.
We consider two types of possible attributes, for primitive/functional parts and for connections. For example, we consider the following attributes, among others, for primitive parts: inertia moments, stability, and regularity, and for connections: ratio of volumes and context based stability. The full description of the geometrical attributes we have considered is relatively large and can be found at [13]
