Abstract
Abstract. This paper presents a method for dense DSM reconstruction from high resolution, mono sensor, passive imagery, spatial panchromatic image sequence. The interest of our approach is four-fold. Firstly, we extend the core of light field approaches using an explicit BRDF model from the Image Synthesis community which is more realistic than the Lambertian model. The chosen model is the Cook-Torrance BRDF which enables us to model rough surfaces with specular effects using specific material parameters. Secondly, we extend light field approaches for non-pinhole sensors and non-rectilinear motion by using a proper geometric transformation on the image sequence. Thirdly, we produce a 3D volume cost embodying all the tested possible heights and filter it using simple methods such as Volume Cost Filtering or variational optimal methods. We have tested our method on a Pleiades image sequence on various locations with dense urban buildings and report encouraging results with respect to classic multi-label methods such as MIC-MAC, or more recent pipelines such as S2P. Last but not least, our method also produces maps of material parameters on the estimated points, allowing us to simplify building classification or road extraction.
Highlights
An extensive body of literature over the subject of dense DSM reconstruction exists, the reconstruction of a reliable DEM or DSM from visible passive optics sensors is still a challenging task nowadays
We recall that the whole image sequence was used in the height estimation
In this study we have provided a way to enlarge light field methods on non-pinhole sensors with non-rectilinear motion using a BRDF model
Summary
An extensive body of literature over the subject of dense DSM reconstruction exists, the reconstruction of a reliable DEM or DSM from visible passive optics sensors is still a challenging task nowadays. The solutions are often a trade-off between pure local radiometric matches (with estimation errors due to the image noise), global priors over the displacement map (varying smoothly, sharp edges, heavy-tail distribution, etc) and the step for the discrete estimation Such problems are sometimes named “multi-label problems”, and graph-cut techniques (Boycov et al, 2001) seemed very promising they only provided an estimation of the sought solution. Improvements were brought in specific cases (Ishikawa, 2003) which could be applied to disparity estimation In this way, the work of (Pock et al, 2008; 2010) seemed even more promising for a global solution was provided, with less memory consumption, in a continuous framework solving a convex problem with higher dimension. The light field method (Kim et al, 2013) casts the disparity estimation problem on all views into a straight line seeking problem, which is very appealing when many views are available
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
