Flow-Based Correspondence Matching in Stereovision

Accurate and efficient correspondence matching between two rectified images is critical for stereo reconstruction. Essentially, correspondence matching co-registers the two rectified images subject to an epipolar constraint (i.e., registration is performed along the horizontal direction). Most algorithms are based on windowed block matching that optimizes cross-correlation or its variants (e.g., sum of squared differences, SSD) between two sub-images to generate a sparse disparity map. In this work, we utilize unrestricted optical flow for a full-field correspondence matching. Relative to surface point measurements sampled with a tracked stylus as ground-truth, we show that the point-to-surface distance from the flow-based method is comparable and often superior to that from the SSD algorithm (e.g., 1.0 mm vs. 1.2 mm, respectively) but with a substantial increase in computational efficiency (5—6 sec for a full field of 41 K vs. 20—30 sec for a sparse subset of 1 K sampling points, respectively). In addition, the flow-based stereovision offers ability for feature identification based on the full-field horizontal disparity map that is directly related to reconstruction pixel depth values, whereas the vertical disparity provides an assessment of the accuracy confidence level in stereo reconstruction, which are not available with SSD methods.