TOWARDS COMPLETE, GEO-REFERENCED 3D MODELS FROM CROWD-SOURCED AMATEUR IMAGES
Keywords: structure from motion, crowd-sourcing, geo-reference
Abstract. Despite a lot of recent research, photogrammetric reconstruction from crowd-sourced imagery is plagued by a number of recurrent problems. (i) The resulting models are chronically incomplete, because even touristic landmarks are photographed mostly from a few “canonical” viewpoints. (ii) Man-made constructions tend to exhibit repetitive structure and rotational symmetries, which lead to gross errors in the 3D reconstruction and aggravate the problem of incomplete reconstruction. (iii) The models are normally not geo-referenced. In this paper, we investigate the possibility of using sparse GNSS geo-tags from digital cameras to address these issues and push the boundaries of crowd-sourced photogrammetry. A small proportion of the images in Internet collections (≈ 10 %) do possess geo-tags. While the individual geo-tags are very inaccurate, they nevertheless can help to address the problems above. By providing approximate geo-reference for partial reconstructions they make it possible to fuse those pieces into more complete models; the capability to fuse partial reconstruction opens up the possibility to be more restrictive in the matching phase and avoid errors due to repetitive structure; and collectively, the redundant set of low-quality geo-tags can provide reasonably accurate absolute geo-reference. We show that even few, noisy geo-tags can help to improve architectural models, compared to puristic structure-from-motion only based on image correspondence.