Man-made world Image Matching over wide Baselines

Abstract

Matching the Images, representing the same content but captured from di erent, widely separated angels, is a fundamental computer vision challenge. SIFT(Scale Invariant Feature Transform) is famous for matching images captured at di erent depths (scales) with little view invariance. SIFT descriptor uses local information around the keypoints to achieve this invariance. In this thesis we added view invariance to these solid SIFT descriptors and used novel views approach to improve matching. We tested our algorithm on di erent categories and found out that matching can be improved by combining local and global descriptors. Our method skips the noise of scene understanding, for example incorrect estimation of 3D box etc., Finally, our approach is more e cient and has potential for real time robotics scenarios.