Geometry-driven Feature Detection

Matching images taken from different viewpoints is a fundamental step for many computer vision applications including 3D reconstruction, scene recognition, virtual reality, robot localization, etc. The typical approaches detect feature keypoints based on local properties to achieve robustness to viewpoint changes, and establish correspondences between keypoints to recover the 3D geometry or determine the similarity between images. The complexity of perspective distortion challenges the detection of viewpoint invariant features; the lack of 3D geometric information about local features makes their matching inefficient.;In this thesis, I explore feature detection based on 3D geometric information for improved projective invariance. The main novel research contributions of this thesis are as follows. First, I give a projective invariant feature detection method that exploits 3D structures recovered from simple stereo matching. By leveraging the rich geometric information of the detected features, I present an efficient 3D matching algorithm to handle large viewpoint changes. Second, I propose a compact high-level feature detector that robustly extracts repetitive structures in urban scenes, which allows efficient wide-baseline matching. I further introduce a novel single-view reconstruction approach to recover the 3D dense geometry of the repetition-based features.