Metric 3D reconstruction from uncalibrated unordered images with hierarchical merging

This paper considers metric 3D reconstruction with hierarchical merging. We first describe pairwise matching of images for spanning tree construction. The epipolar constraint is imposed on two related vertices. We then present a partition method to group the images into duplets or triplets for partial reconstruction. A robust hierarchical scheme that merges partial reconstruction together is used to form a complete projective one. Finally, we describe a robust self-calibration method, allowing for more reliable upgrade to metric geometry. The proposed scheme could handle uncalibrated unordered images and was tested on real images. Experimental results show that it can give correct scene structure and camera motion across multiple widely separated images.