Learning Informative SIFT Descriptors for Attentive Object Recognition

Abstract

With the emerging sensor technologies in mobile devices, such as cameraphones, visual interpretation methodologies are challenged to providesolutions within the everydays outdoor urban environment. For thispurpose, we propose to apply the 'Informative Descriptor Approach'on the SIFT descriptor [4], in order to define the informative SIFT(i-SIFT) descriptor. By attentive matching of i-SIFT keypoints, weprovide an innovative method on object detection that significantlyimproves SIFT based keypoint matching. i-SIFT tackles the SIFT bottlenecks,e.g., extensive nearest neighbor indexing, by (i) significantly reducingthe descriptor dimensionality, (ii) decreasing the size of objectrepresentation by one order of magnitude, and (iii) performing matchingexclusively on attended descriptors, as required by resource sensitivedevices. The key advantages of informative SIFT (i-SIFT) are demonstratedin a typical outdoor mobile vision experiment on the TSG-20 referencedatabase, detecting buildings with high accuracy.