Learning Informative SIFT Descriptors for Attentive Object Detection

With the emerging sensor technologies in mobile devices, such as camera phones, visual interpretation methodologies are challenged to provide solutions within the everydays outdoor urban environment. For this purpose, we propose to apply the 'Informative Descriptor Approach' on the SIFT descriptor [4], in order to de ne the informative SIFT (i-SIFT) descriptor. By attentive matching of i-SIFT keypoints, we provide an innovative method on object detection that signifantly improves SIFT based keypoint matching. i-SIFT tackles the SIFT bottlenecks, e.g., extensive nearest neighbor indexing, by (i) signifantly reducing the descriptor dimensionality, (ii) decreasing the size of object representation by one order of magnitude, and (iii) performing matching exclusively on attended descriptors, as required by resource sensitive devices. The key advantages of informative SIFT (i-SIFT) are demonstrated in a typical outdoor mobile vision experiment on the TSG-20 reference database, detecting buildings with high accuracy.