Abstract
Abstract. Several 3D reconstruction pipelines are being developed around the world for satellite imagery. Most of them implement their own versions of Semi-Global Matching, as an option for the matching step. However, deep learning based solutions already outperform every SGM derived algorithms on Kitti and Middlebury stereo datasets. But these deep learning based solutions need huge quantities of ground truths for training. This implies that the generation of ground truth stereo datasets, from satellite imagery and lidar, seems to be of great interest for the scientific community. It will aim at reducing the potential transfer learning difficulties, that could arise from a training done on datasets such as Middlebury or Kitti. In this work, we present a new ground truth generation pipeline. It produces stereo-rectified images and ground truth disparity maps, from satellite imagery and lidar. We also assess the rectification and the disparity accuracies of these outputs. We finally train a deep learning network on our preliminary ground truth dataset.
Highlights
Several 3D reconstruction pipelines are being developed around the world for satellite imagery
In the frame of the CNES / Airbus CO3D mission (Lebegue et al, 2020), CNES, the French space agency, is developing its own pipeline, called CARS (Youssefi et al, 2020). This new multi-view stereo pipeline is focused on robustness and scalability, as it will be used for massive Digital Surface Model (DSM) production (Melet et al, 2020)
Deep learning based solutions already outperform every SGM derived algorithms on Kitti and Middlebury stereo datasets (Menze et al, 2018) (Scharstein, Szeliski, 2002). It means that multi-view stereo pipelines users could be interested in testing the most promising deep-learning approaches for the stereo-matching step
Summary
Several 3D reconstruction pipelines are being developed around the world for satellite imagery. In the frame of the CNES / Airbus CO3D mission (Lebegue et al, 2020), CNES, the French space agency, is developing its own pipeline, called CARS (Youssefi et al, 2020) This new multi-view stereo pipeline is focused on robustness and scalability, as it will be used for massive DSM production (Melet et al, 2020). Deep learning based solutions already outperform every SGM derived algorithms on Kitti and Middlebury stereo datasets (Menze et al, 2018) (Scharstein, Szeliski, 2002) It means that multi-view stereo pipelines users could be interested in testing the most promising deep-learning approaches for the stereo-matching step (just replacing SGM based solutions by these ones). Pandora will be publicly available as an open-source software; and it will be the stereo-matching tool of the future CO3D mission ground segment
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have