The Taub Faculty of Computer Science Events and Talks

Correctly matching feature points across images is an important preprocessing step for many computer vision applications (specifically such that require geometric reasoning). Once an initial set of putative matches is obtained, the common methods to detect the correct matches, e.g., RANSAC, use a parametric model that relates the correct matches between an image pair, e.g., the fundamental matrix. To obtain an accurate (uncontaminated) such model, a large number of search iterations is usually required, which results in a relatively long running time (a few second for an image pair). This becomes a bottleneck (running-time wise) when considering thousands of image pairs as in a SfM pipeline.

In this talk we will introduce a weaker (but simpler) non-parametric geometric model that takes into account only the spatial order between features in images. We will discuss applications that can benefit from this model; in particular, a novel application to guide a camera’s rotation such that it can capture a specified region of interest in the scene. We will further show that this model allows to quickly estimate the number of correct matches between an image pair without directly computing them. The number of correct matches will then be shown to be useful in several applications. For example, skipping the estimation of fundamental matrices of image pairs with small number of correct matches in a SfM pipeline, resulting in a reduction of ~%99 of the running time while preserving most of the accuracy (~80% of the correct matches).

Joint work with Yael Moses and Ilan Shimshoni.