Interest Point Detectors and Descriptors for IR Images : An Evaluation of Common Detectors and Descriptors on IR images

Abstract: Interest point detectors and descriptors are the basis of many applications within computer vision. In the selection of which methods to use in an application, it is of great interest to know their performance against possible changes to the appearance of the content in an image. Many studies have been completed in the field on visual images while the performance on infrared images is not as charted. This degree project, conducted at FLIR Systems, provides a performance evaluation of detectors and descriptors on infrared images. Three evaluations steps are performed. The first evaluates the performance of detectors; the second descriptors; and the third combinations of detectors and descriptors. We find that best performance is obtained by Hessian-Affine with LIOP and the binary combination of ORB detector and BRISK descriptor to be a good alternative with comparable results but with increased computational efficiency by two orders of magnitude.