Incremental Reconstruction of Manifold Surface from Sparse Visual Mapping

Abstract

International audienceAutomatic image-based-modeling usually has two steps: Structure from Motion (SfM) and the estimation of a trian-gulated surface. The former provides camera poses and a sparse point cloud. The latter usually involves dense stereo. From the computational standpoint, it would be nice to avoid dense stereo and estimate the surface from the sparse cloud directly. Furthermore, it would be useful for online applications to update the surface while the camera is moving in the scene. This paper deals with both requirements: it introduces an incremental method which reconstructs a surface from a sparse cloud estimated by incremental SfM. The context is new and difficult since we ensure the resulting surface to be manifold at all times. The manifold property is important since it is needed by differential operators involved in surface refinements. We have experimented with a hand-held omnidirectional camera moving in a city

Incremental Solid Modeling from Sparse and Omnidirectional Structure-from-Motion Data

“…Several sculpting methods [5,13,15,18,20] use visibility knowledge and sparse point cloud P estimated from images. The manifold constraint on |δV | is not enforced by [13,15,18].…”

“…In [13], a non-incremental method estimates a surface minimizing a cost involving visibility and photoconsistency. The manifold constraint is enforced in [5,20]. Method [5] is limited to genus zero surface.…”

“…Method [5] is limited to genus zero surface. Method [20] is incremental but the time complexity of one iteration can be too large: if the camera trajectory is a loop, this complexity is at least linear to the number of camera poses in the loop.…”

Incremental Surface Extraction from Sparse Structure-from-Motion Point Clouds

“…Lovi et al [14] extend the space carving approach to work in an incremental manner but their method requires additional memory that grows with the scene size. Lhuillier et al [19] present a method that incrementally extracts a surface from sparse points generated by a visual odometry system. Instead of carving free-space tetrahedra, they greedily aggregate free-space tetrahedra and the outer envelope defines the surface.…”

“…To sum up, the aforementioned approaches for incremental surface reconstruction are either based on a strong camera motion assumption [19] or they are not robust to outliers [14,15]. By contrast, robust methods like [13] are not well suited to be implemented in an incremental manner as we show later.…”