| For designing detectors for infrequently occurring objects in wide-area satellite imagery, we are faced with the challenge that such objects are difficult to find for the purpose of generating the needed training data. As a result, a human agent must expend an inordinate length of time in order to produce a sufficient number of labeled training data. In this dissertation, we reduce this annotation burden by drawing upon the research that has been carried out recently in the area of active learning, whereby the machine searches for human annotation those unlabeled samples that can actually improve the detector. The search is iterative: Starting with a small number of human-supplied strongly positive and negative samples, our framework scans the images and tests the candidate samples against the current decision surfaces. Only those samples that are too close to the decision surfaces are sent to the human for annotation and the new samples thus labeled used to update the decision surfaces. We have applied this framework to create detectors for pedestrian crosswalks and transmission-line towers in a cloud-based implementation in areas of over 150,000 sq km in Australia. We should also mention that a stepping stone to this work was our earlier, more direct approach to detector design in which the class discrimination features are specified by the human. This approach, which was also an exercise in exploiting volunteer-generated road maps for detecting objects that are on or along the roads, was used to create a crosswalk detector and applied to a 180,000 sq km area in Australia. |
