Ellipse Constraints for Improved Wide-Baseline Feature Matching and Reconstruction

The classic feature matching process has two drawbacks. Firstly, ambiguous but possibly correct matches will potentially be removed and secondly, there is no constraint for the 2D size of the features. In the present paper these drawbacks are tackled at once with a different approach: by considering region features instead of point features and by adding constraints based on the features' shape. Here, the shape will be described with an ellipse. Using existing knowledge about the algebraic properties of ellipses within the computer vision domain, this enables additional constraints such as ellipse tangents. The number of ambiguous matches is reduced and increased control of the physical 2D size of the features is obtained. This will be shown on known epipolar geometry. Additionally, reconstruction of feature ellipses is examined.