Feature matching is one of the most important vision problems. This page demonstrates a novel algorithm for determining the correspondence between two sets of coplanar points and lines. The algorithm is based solely on geometric information, namely the two-line two-point projective invariant and the plane homography that relates two views of the same 3D plane. The algorithm is robust to the existence of outliers, i.e. features that do not have matching counterparts in one of the two images. Outliers are introduced by various sources, such as errors during feature extraction, occlusions, existence of non-planar objects, etc. Four experiments performed with a prototype implementation of the algorithm are described below. A paper describing the algorithm appears in the proceedings of BMVC'98.

The first experiment employs a pair of views of a poster lying on the floor. The extracted points and lines are shown here in red and green respectively. Matched features are shown here with identical labels. The two views can be registered by warping the second view according to the plane homography estimated from the matched features.

The second experiment is based on a pair of aerial views of an approximately planar surface. The extracted points and lines are shown here in red and green respectively. Notice that the two views contain a large amount of outliers. Matched features are shown here with identical labels. The two views can be registered by warping the second view according to the plane homography estimated from the matched features.

The third experiment is based on the well-known "pentagon'' stereo pair. To make the experiment more difficult, the right image has been manualy rotated by 95 degrees counterclockwise. The extracted points and lines are shown here in red and green respectively. Although the scene is not exactly planar, it can be considered as such due to its large distance from the camera. Matched features are shown here with identical labels. The two views can be registered by warping the right view according to the plane homography estimated from the matched features.

The last experiment applies the algorithm on a series of images depicting a wall of a room. The left and right image of each row were in turn matched with the middle one and then the middle images of the first and third row were matched with the middle image of the second row. By employing the estimated homographies, each image was warped towards the middle image of the second row, so as to construct a mosaic. This mosaic corresponds to the appearance of the wall from the viewpoint of the middle image of the second row.