Building detection from mobile imagery using informative SIFT descriptors

We propose reliable outdoor object detection on mobile phone imagery from off-the-shelf devices. With the goal to provide both robust object detection and reduction of computational complexity for situated interpretation of urban imagery, we propose to apply the 'Informative Descriptor Approach' on SIFT features (i-SIFT descriptors). We learn an attentive matching of i-SIFT keypoints, resulting in a significant improvement of state-of-the-art SIFT descriptor based keypoint matching. In the off-line learning stage, firstly, standard SIFT responses are evaluated using an information theoretic quality criterion with respect to object semantics, rejecting features with insufficient conditional entropy measure, producing both sparse and discriminative object representations. Secondly, we learn a decision tree from the training data set that maps SIFT descriptors to entropy values. The key advantages of informative SIFT (i-SIFT) to standard SIFT encoding are argued from observations on performance complexity, and demonstrated in a typical outdoor mobile vision experiment on the MPG-20 reference database.