Challenges of multi-view satellite stereo reconstruction pipelines and some contributions on key stages.
Satellite imagery is quickly gaining in importance, with Earth observation satellites producing daily images from all the points of the globe, both commercially and freely available. In this thesis we concentrate on surface reconstruction from visible light satellite images through stereo-vision. Gi...
Gorde:
| Egile nagusia: | |
|---|---|
| Formatua: | doctoralThesis |
| Hizkuntza: | ingelesa |
| Argitaratua: |
2023
|
| Gaiak: | |
| Sarrera elektronikoa: | https://hdl.handle.net/20.500.12008/36993 |
| Etiketak: |
Etiketarik gabe, Izan zaitez lehena erregistro honi etiketa jartzen!
|
| Gaia: | Satellite imagery is quickly gaining in importance, with Earth observation satellites producing daily images from all the points of the globe, both commercially and freely available. In this thesis we concentrate on surface reconstruction from visible light satellite images through stereo-vision. Given two images of a scene from different known viewpoints, the objective of stereo is to estimate the most likely 3D shape or depth that explains those images. When more than two images are available, multi-view stereo (MVS) can be applied working by pairs and integrating the reconstructions (pair-wise MVS) or deriving a reconstruction from all the images at a time (true MVS). In the case of satellite images, MVS has traditionally been performed with pair-wise approaches where the multiple views are treated by pairs doing traditional two-view stereo and then aggregating the digital surface models (DSM) from the pair-wise reconstructions to get the final result. Several well established commercial and open-source solutions organize their working pipelines in this way. This solutions mostly rely on classic stereo algorithms while deep learning (DL) alternatives are slowly being adapted to work in the pipelines. But the DL based approaches have not still clearly outperformed the traditional pipelines and there is room for much more work in this yet open area. A crucial issue that complicates the advance in this field is the scarce public datasets with well curated ground-truth. In this thesis a set of methods from different approaches of pair-wise and true MVS were evaluated and compared. For the comparison, classic and deep learning methods were adapted to work with satellite images and to correctly interface with S2P, a modular satellite stereo pipeline. The results obtained with deep learning methods showed the potential of using this kind of algorithms on satellite images as a step in a classic pipeline or as an end-to-end MVS solution. Considering pair-wise MVS, besides the stereo matching, two other steps are crucial to achieve a good reconstruction: (a) the selection of the most appropriate pairs, and (b) the fusion of the DSMs reconstructed from the pairs. For pair selection, a novel strategy based on the simulation of satellite images was devised and can order the pairs in a more consistent way than commonly used heuristics. For the simulation of images, a tool that can generate views from an artificial 3D scene was developed. Regarding the fusion of DSMs, an iterative scheme based on the bilateral filtering was conceived showing to be a robust and performant method. Improvements in other stages of the baseline stereo pipeline and the processing and analysis of point clouds were also part of the topics addressed during the thesis. |
|---|