adaptive context-dependent classification

alexandre r.j. françois

synopsis || acdc || color classification

synopsis

This project proposes a novel, dynamic computational model for adaptive context-dependent classification (AC-DC).

the ac-dc model

AC-DC addresses online categorization of samples of a signal according to classes for which a distance or similarity is defined. AC-DC operates under the assumption that the input signal has some underlying structure such that explanation (classification) in the context of what was usually produces a reasonable approximation for what is. Given time, a more accurate explanation might become available through an adaptaion mechanism. when a satisfactory explanation is not known, one might be created through a learning process. The dynamic nature of the process emphasizes the balance between the time required (or available) to produce a classification solution, and the accuracy of the solution.

The model consists of three subsystems that operate concurrently:

• The classification subsystem determines, for each incoming sample, the best corresponding class in the current context (subset of all known classes).
• The adaptation subsystem regulates the movement of the known classes between the current context and one or several other set of classes of out-of-context classes, based on their estimated relevance to the explanation of current and recent incoming samples.
• The learning subsystem handles the creation of new classes when no existing class can satisfactorily explain a sample. Another long term learning mechanism, more pervasive, regulates the adaptation of known classes in response to the input signal.

The AC-DC model comprises of 3 concurrent subsystems: classification, context adaptation, and learning.

adaptive context-dependent color classification

19 September 2008

This demonstration showcases the AC-DC model in an application to color classification. A real-time, concurrent software system, designed and implemented using the SAI architectural framework, affords interactive illustration and experimentation on a laptop with webcam. [mov]

The demonstration system applies AC-DC to color classification. In uncompressed video, pixel color values can span a large, uniformly and finely sampled space. For applications such as video compression and analysis, it is often desirable to group pixels whose color values are similar into color classes. The dynamic and adaptive nature of the AC-DC model sets it apart from other methods traditionally applied to such problems.

The input stream consists of live video encoded in the YUV color model. The classification stage assigns to each pixel a color label. the label is used in color segmentation to group pixels with the same label into color regions. The distance of the original pixel value to the color class assigned to the pixel is a measure of how good an explanation of the pixel color value the current context affords, and therefore provides an estimation of novelty.

The real-time software system, designed in the SAI architectural style and implemented using the MFSM open source architectural middleware, runs on a laptop with a webcam (Mac Book Pro with built-in iSight in this case). The system performs video capture, color classification, context adaptation, learning, result visualization (color regions or novelty map) computation and display. The system also displays a graphical rendering of the color classes in the different levels of the classification stack. The system design is inherently concurrent and asynchronous.

Users interact with the system by moving in front of the camera and/or presenting objects of various colors. They can observe the effect on the dynamic system in real-time. Users can also change the parameters that influence the dynamic properties of the system and observe the resulting change in behavior.