Semi-Supervised Learning with Visual Pixel-Level Similaries for Object Detection

We introduce a novel approach for detection of objects from aerial images at the level of pixels using semi-supervised learning. Buildings in aerial images are complex 3D objects which are represented by features of different modalities include visual information and 3D height data. Semi-supervised learning is a classification which additional unlabeled data can be used to improve accuracy. This aims to use semi-supervised boosting learning offer an interesting solution to this problem by learning from both labeled and unlabeled data. The major advantage of this approach is the simplicity with which the prior knowledge is incorporated into the semi-supervised learning mechanism. We demonstrate an early result of using semi-supervised boosting learning algorithm to indeed detect building in aerial digital imagery to a satisfactory and useful level of completeness.