Robust extraction of image correspondences exploiting the image scene geometry and approximate camera orientation

Image-based modeling techniques are an important tool for producing 3D models in a practical and cost effective manner. Accurate image-based models can be created as long as one can retrieve precise image calibration and orientation information which is nowadays performed automatically in computer vision and photogrammetry. The first step for orientation is to have sufficient correspondences across the captured images. Keypoint descriptors like SIFT or SURF are a successful approach for finding these correspondences. The extraction of precise image correspondences is crucial for the subsequent image orientation and image matching steps. Indeed there are still many challenges especially with wide-baseline image configuration. After the extraction of a sufficient and reliable set of image correspondences, a bundle adjustment is used to retrieve the image orientation parameters. In this paper, a brief description of our previous work on automatic camera network design is initially reported. This semi-automatic procedure results in wide-baseline high resolution images covering an object of interest, and including approximations of image orientations, a rough 3D object geometry and a matching matrix indicating for each image its matching mates. The main part of this paper will describe the subsequent image matching where the pre-knowledge on the image orientations and the pre-created rough 3D model of the study object is exploited. Ultimately the matching information retrieved during that step will be used for a precise bundle block adjustment. Since we defined the initial image orientation in the design of the network, we can compute the matching matrix prior to image matching of high resolution images. For each image involved in several pairs that is defined in the matching matrix, we detect the corners or keypoints and then transform them into the matching images by using the designed orientation and initial 3D model. Moreover, a window is defined for each corner and its initial correspondence in the matching images. A SIFT or SURF matching is implemented between every matching window to find the homologous points. This is followed by Least Square Matching LSM to refine the correspondences for a sub-pixel localization and to avoid inaccurate matches. Image matching is followed by a bundle adjustment to orient the images automatically to finally have a sparse 3D model. We used the commercial software Photomodeler Scanner 2010 for implementing the bundle adjustment since it reports a number of accuracy indices which are necessary for the evaluation purposes. The experimental test of comparing the automated image matching of four pre-designed streopairs shows that our approach can provide a high accuracy and effective orientation when compared to the results of commercial and open source software which does not exploit the pre-knowledge about the scene.