PhD Thesis On-line Conservative Learning

I n this thesis we are concerned with classifier-based object detection. Therefore, a previously learned classifier is iteratively applied on all possible positions in a given image. When learning an object representation one faces two main problems. First, a representative set of training samples is required, which is usually obtained by hand labeling. Since for complex object classes this set may be quite large, this is not convenient. Second, due to the high variability of an object’s appearance and varying environmental conditions general object representations often fail in practice (low recall, low precision). Hence, it would be beneficial to have an adaptive system. But such systems have to learn from unknown, unlabeled samples. Thus, in this thesis we propose a method for visual learning, that addresses both problems. In fact, Conservative Learning minimizes the manual (labeling) effort when on-line learning an adaptive classifier. This is realized by combining the power of a discriminative model and the robustness of a generative model. The main idea is to verify the detection results of the discriminative model by the generative model and to classify those patches either as true positive or as false positive. Since we have a huge amount of data (i.e., a video stream) we can use very conservative parameters for this decision. Hence, most samples are not considered at all. The thus labeled samples are fed back to the system and we get an increasingly better detector. To get the whole process started we acquire training samples from low level cues such as background subtraction, motion detection, or tracking. These cues can also be applied to acquire additional positive samples. Experimental evaluations demonstrate the benefits of the approach. In fact, we analyze the learning behaviour and show that an increasingly better detector is obtained over time. Additionally, the approach is demonstrated for various different applications including detecting persons, faces, and hand held objects. For the person detection task, that is analyzed more detailed, we show that our approach outperforms state-of-the-art methods, that require a time-consuming off-line training.