Dynamic object localization using hand-held cameras

Abstract

We consider the problem of separating static and dynamic regions of a scene when the camera also undergoes motion while capturing still images. We assume that we do not have any other information about the scene and the camera settings. Given two images, we would like to estimate the static and dynamic regions corresponding to one image with respect to the other image. The proposed solution involves over-segmentation of the image and dense correspondence between the two images. We show that the proposed approach works well even when there are changes in illumination and exposure settings while capturing the two images. We evaluate the performance of the proposed approach by demonstrating the results on different scenes with complex object motions.