Visual sensor networks have emerged as an important class of sensor-based distributed intelligent systems, where image matching is one of the key technologies. This article presents an affine invariant method to produce dense correspondences between uncalibrated wide baseline images. Under affine transformations, both point location and its neighborhood texture are changed between views, so dense matching becomes a tough task. The proposed approach tends to solve this problem within a sparse-to-dense framework. The contribution of this article is in threefolds. First, a strategy of reliable sparse matching is proposed, which starts from affine invariant features extraction and matching and then these initial matches are utilized as spatial prior to produce more sparse matches. Second, match propagation from sparse feature points to its neighboring pixels is conducted in the way of region growing in an affine invariant framework. Third, the unmatched points are handled by low-rank matrix recovery technique. Comparison experiments of the proposed method versus existing ones show a significant improvement in the presence of large affine deformations.
MBR-SIFT: A mirror reflected invariant feature descriptor using a binary representation for image matching
