Boosting in location space

Computer-based object detection promises to vastly change our lives. Robots will be able to map their environment and make sense of the world. Scientists will have a new pair of eyes to sift through terabytes of images of molecules, proteins, waterways, and galaxies to discover novel science. Most approaches to object detection use general-purpose machine learning algorithms that optimize a non-spatial objective. The classic sliding window approach predicts the presence or absence of an object in every window of an image; this optimizes the classifier for detection but not for localization. In contrast, this thesis introduces a new boosting algorithm (LocationBoost) that operates over an entire image at all times during training, directly predicts object locations, and minimizes a spatial loss function that is strongly motivated by object detection. The research of this dissertation led to six major contributions in object detection. First, it is impossible to develop a universal metric to evaluate the accuracy on any object detector on every object detection task. We do not exclusively use a single “standard” metric. Instead, we clearly define three different problems in small object detection and devise metrics well-matched to them. Second, we introduce generative grammars to combine primitive image features into composite features. Composite features are more informative and lead to more accurate object detection. Third, AdaBoost is fragile in the presence of noisy and ambiguous training data but we made spatially exploitative adaptations to the learning algorithm to greatly improve learning stability. Fourth, a radically different boosting algorithm (LocationBoost ) is proposed that directly locates centers of small objects, bypassing the need for bounding boxes. Instead of boosting classifiers that predict whether or not a patch contains an object, our new approach boosts object detectors that produce a list of predicted object locations. LocationBoost uses a new spatial loss function reflecting the intuition that large areas of background are uninteresting and not worth spending computational effort on. Fifth, LocationBoost is extended so it can predict bounding boxes that enclose objects. Sixth, a multi-scale variant of LocationBoost is proposed to enable the detection of both large and small objects in the same image. In this variant, we show how the structure of multi-scale detection can be exploited to greatly speed up training and detection. Seventh, we propose a new primitive image feature based on FAST corner detection that enables real-time object detection.