Towards Efficient 3D Reconstructions from High-Resolution Satellite Imagery Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 21, 2019
  • Wang, Ke
    • Affiliation: College of Arts and Sciences, Department of Computer Science
  • Recent years have witnessed the rapid growth of commercial satellite imagery. Compared with other imaging products, such as aerial or streetview imagery, modern satellite images are captured at high resolution and with multiple spectral bands, thus provide unique viewing angles, global coverage, and frequent updates of the Earth surfaces. With automated processing and intelligent analysis algorithms, satellite images can enable global-scale 3D modeling applications. This dissertation explores computer vision algorithms to reconstruct 3D models from satellite images at different levels: geometric, semantic, and parametric reconstructions. However, reconstructing satellite imagery is particularly challenging for the following reasons: 1) Satellite images typically contain an enormous amount of raw pixels. Efficient algorithms are needed to minimize the substantial computational burden. 2) The ground sampling distances of satellite images are comparatively low. Visual entities, such as buildings, appear visually small and cluttered, thus posing difficulties for 3D modeling. 3) Satellite images usually have complex camera models and inaccurate vendor-provided camera calibrations. Rational polynomial coefficients (RPC) camera models, although widely used, need to be appropriately handled to ensure high-quality reconstructions. To obtain geometric reconstructions efficiently, we propose an edge-aware interpolation-based algorithm to obtain 3D point clouds from satellite image pairs. Initial 2D pixel matches are first established and triangulated to compensate the RPC calibration errors. Noisy dense correspondences can then be estimated by interpolating the inlier matches in an edge-aware manner. After refining the correspondence map with a fast bilateral solver, we can obtain dense 3D point clouds via triangulation. Pixel-wise semantic classification results for satellite images are usually noisy due to the negligence of spatial neighborhood information. Thus, we propose to aggregate multiple corresponding observations of the same 3D point to obtain high-quality semantic models. Instead of just leveraging geometric reconstructions to provide such correspondences, we formulate geometric modeling and semantic reasoning in a joint Markov Random Field (MRF) model. Our experiments show that both tasks can benefit from the joint inference. Finally, we propose a novel deep learning based approach to perform single-view parametric reconstructions from satellite imagery. By parametrizing buildings as 3D cuboids, our method simultaneously localizes building instances visible in the image and estimates their corresponding cuboid models. Aerial LiDAR and vectorized GIS maps are utilized as supervision. Our network upsamples CNN features to detect small but cluttered building instances. In addition, we estimate building contours through a separate fully convolutional network to avoid overlapping building cuboids.
Date of publication
Resource type
  • Niethammer, Marc
  • Berg, Alexander
  • Manocha, Dinesh
  • Dunn, Enrique
  • Frahm, Jan-Michael
  • Doctor of Philosophy
Degree granting institution
  • University of North Carolina at Chapel Hill
Graduation year
  • 2018

This work has no parents.