Using Normalized Interest Point Trajectories Over Scale for Image Search

Abstract

Image search and object recognition are two domains where it is useful to be able to describe an image in a form that is invariant to image lighting, image intensity, scaling, rotation, translation, and changes in camera position. This paper presents a method based on tracing the trajectories of interest points, specifically KLT corners, across scale-space. The KLT corner interest points are calculated with an adaptive threshold to make them invariant to image intensity. A three-dimensional point composed of two-dimensional spatial coordinates and the scale of gaussian smoothing is found for each interest point, together all the points in the image are normalized into a form that is mostly invariant to geometric changes such as scale and rotation. Each image is converted to a trajectory set which is compared between images to assess their similarity. Experiments are shown.