To Bundle Adjust or Not: A Comparison of Relative Geolocation Correction Strategies for Satellite Multi-View Stereo

Abstract

The generation of up-to-date accurate 3D models from multi-view satellite images has recently become a hot research topic. A well-known challenge of this problem is to put all cameras into a common frame of reference, since depending on the satellite geopositioning equipment the camera parameters may contain errors of up to tens of meters on the ground. In this context, bundle adjustment based techniques, relying on the identification of a set of tie-points and the correction of the camera models to make them coincident, have become a generally accepted practice. However, new approaches capable of producing state-of-the-art results without the use of prior bundle adjustment have also been proposed. This work aims to compare both strategies and assess the practical impact of using bundle adjustment for 3D reconstruction from multi-view satellite images.

State of the Art in Digital Surface Modelling From Multi-View High-Resolution Satellite Images

“…However, it depends on detecting and matching a sufficient quantity of accurate correspondences across the multiview images, which is a problem when the study area is restricted to a small test area of interest. Moreover, previous work (Marí et al, 2019) highlighted that post-co-registration of the computed multi-stereo DSMs showed high robustness and accuracy compared to the bundle adjustment for small area scenario. Hence in this paper, no bundle adjustment is conducted before multi-stereo matching;…”

A Generic Bundle Adjustment Methodology for Indirect RPC Model Refinement of Satellite Imagery

“…Mismatches in the feature track observations can strongly undermine the bundle adjustment performance, leading to failure or strong biases [3,29]. We remove erroneous matches by means of a distance ratio test with a relative threshold of 0.6 [28].…”

“…Similarly to [29], the object coordinates of the tie point that projects onto each feature track are initialized by triangulating all the pairwise matches of the track and taking the mean of the resulting 3D locations (Algorithm 3). The RPC triangulation algorithm from [15], denoted by Triangulate in Algorithm 3, is employed to triangulate keypoint correspondences with the RPC camera models.…”

“…where φ m , θ m and α m are the three Euler angles to optimize for the m-th camera. This formulation is similar to the one introduced in [29], the key distinction is that in this work the estimated rotation corrects the original RPC model itself instead of a local affine camera that approximates it.…”

“…with a point cloud registration algorithm. Previous works show that geometry based alignments can be used at a local scale to elude any prior RPC correction [18,24,29], but they offer poor scalability as the amount of models to register increases. Running a bundle adjustment to refine the RPC models seems a more natural and less-constrained solution to the problem, as long as enough tie points are available.…”