Learning and Detection of Higher-Level Primitives Using AdaBoost

In the task of facade interpretation higher-level interpretation system has to cooperate with lower-level image processing modules (IPMs). The AdaBoost IPM was designed to work as a lower-level image interpretation module, working directly with images. Its outputs are confidence-rated hypotheses of positions of objects of interest in the image. A higher-level reasoning module (e.g. SCENIC) is expected to run the module, use its outputs for further reasoning and send “down” feedback on both learning and classification results of the IPM. This process can be repeated (reasoning loop) until satisfactory scene interpretation is obtained. Several facade primitives like T-style windows and triangular cornices were tested as exemplar objects of interest. The main advantage of the AdaBoost IPM is its scalability in object types and extensibility to the online training and classifier refinement.

References