Machine Learning for Image Analysis

This project focused on applying machine learning techniques to problems in machine vision. Our early efforts focused on adapting learning methods to acquire three-dimensional models of object classes from training images. In more recent work, we evaluated a variety of induction techniques (e.g., CN2, C4.5, nearest neighbor, naive Bayes, perceptron learning) on their ability to aid detection of rooftops in aerial photographs. Our training and test data were provided by Ram Nevatia and his colleagues in the Institute for Robotics and Intelligent Systems at the University of Southern California.

The detection of rooftops is one of many levels of processing in the USC-developed BUDDS vision system. At each level of processing, it vision system posits a set of hypotheses (e.g., roof hypotheses), which are accepted (i.e., roofs) or rejected (i.e., non-roofs) by a selection process. The implementation we were attempting to improve upon used a hand-configured linear classifier to select and reject roof hypotheses. Our research hypothesis was that machine learning algorithms could provide a faster method for configuring classifiers for this task and that the accuracy of the machine-learned classifiers will be equivalent to, if not better than, the accuracy of the hand-coded classifiers.

This work was funded by the Image Understanding Program of the Defense Advanced Research Projects Agency, through Grant N00014-94-1-0746 from the Office of Naval Research, and through an equiment grant from Sun Microsystems.

One of the aerial photographs used in our experiments on learning to detect rooftops.

Related Publications

Maloof, M. A., Langley, P., Binford, T. O., Nevatia, R., & Sage, S. (2003). Improved rooftop detection in aerial images with machine learning. Machine Learning, 53, 157-191.
Ali, K. M., Langley, P., Maloof, M. A., Binford, T. O., & Sage, S. (1998). Improving rooftop detection with interactive visual learning. Proceedings of the Image Understanding Workshop. Monterrey, CA: Morgan Kaufmann.
Maloof, M. A., Langley, P., Binford, T. O., & Nevatia, R. (1998). Generalizing over aspect and location for rooftop detection. Proceedings of the Fourth IEEE Workshop on Applications of Computer Vision. Princeton, NJ: IEEE Press.
Maloof, M. A., Langley, P., Binford, T., & Sage, S. (1998). Improving rooftop detection in aerial images through machine learning (Technical Report 98-1). Institute for the Study of Learning and Expertise, Palo Alto, CA.
Maloof, M. A., Langley, P., Sage, S., & Binford, T. (1997). Learning to detect rooftops in aerial images. Proceedings of the 1997 Image Understanding Workshop (pp. 835-845). New Orleans: Morgan Kaufmann.
Provan, G., Langley, P., & Binford, T.O. (1996). Probabilistic learning of three-dimensional object models. Proceedings of the Image Understanding Workshop (pp. 1403-1413). Palm Springs, CA: Morgan Kaufmann.
Langley, P., Binford, T. O., & Levitt, T. S. (1994). Learning object models from visual observation and background knowledge. Proceedings of the Image Understanding Workshop. Monterrey, CA.
Bowyer, K. W., Hall, L. O., Langley, P., Bhanu, B., & Draper, B. A. (1994). Report of the AAAI fall symposium on machine learning and computer vision: What, why and how. Proceedings of the Image Understanding Workshop. Monterrey, CA.

For more information, send electronic mail to langley@isle.org