A saliency-based framework for 2D-3D registration

Here we propose a saliency-based filtering approach to the problem of registering an untextured 3D object to a single monocular image. The principle of saliency can be applied to a range of modalities and domains to find intrinsically descriptive entities from amongst detected entities, making it a rigorous approach to multi-modal registration. We build on the Kadir-Brady saliency framework due to its principled information-theoretic approach which enables us to naturally extend it to the 3D domain. The salient points from each domain are initially aligned using the SoftPosit algorithm. This is subsequently refined by aligning the silhouette with contours extracted from the image. Whereas other point based registration algorithms focus on corners or straight lines, our saliency-based approach is more general as it is more widely applicable e.g. to curved surfaces where a corner detector would fail. We compare our salient point detector to the Harris corner and SIFT keypoint detectors and show it generally achieves superior registration accuracy.