Abstract

Beamer is a new system for unstructured object detection: from an input image, it emits a list of (x,y) pairs, which are the predicted locations of objects. The system shows excellent results in the presence of noisy and ambiguous greyscale aerial imagery, and I describe key elements of the approach to achieve these results.

First, Beamer incorporates domain expertise by generating a broad set of useful features with a stochastic generative grammar, which also eliminates many unhelpful features. A pixel classification is represented with a weighted linear combination of these features, which is learned with the popular boosting algorithm AdaBoost. Second, to improve robustness to label and data noise without impeding localization performance, we soften the learning goal of correctly classifying every pixel in an object by optimizing a post-processing filter applied to weak pixel classifications, which makes use of confidence-rated AdaBoost. Since a purely pixel-based approach to learning is limited for predicting locations, the final detector is optimized with (x,y) location-space criteria.

Grammar features are broadly applicable to many computer vision problems so I will give a mini-tutorial on how to use them in your own computer vision research. GGFE (http://ggfe.googlecode.com) is a new open source library written in Python for this purpose. I describe the basics of generative grammars, generating programs from them, and finally how to incorporate image processing algorithms into a grammar.

Finally, I conclude with a comprehensive experimental evaluation where I also show that the performance of an object detector greatly depends on how one defines performance. By changing the criteria, I show how the ROC curves change considerably. I define three real-world problems and two well-matched criteria for optimizing these problems to evaluate Beamer. The results show all layers of the system greatly boost accuracy.