Method for 3D reconstruction from satellite imagery

1. A method for 3D reconstruction from satellite imagery using deep learning, said method comprising the steps of:

providing ( 101 ) at least two overlapping 2D satellite images,

providing ( 102 ) imaging device parameters for the at least two overlapping 2D satellite images,

providing ( 103 ) at least one trained Machine Learning Network, MLN, able to predict depth maps, said trained MLN being trained on a training set comprising multi-view 3D geocoded ground truth data, and

predicting ( 104 ) a depth map of the at provided at least two 2D satellite images using the trained at least one MLN and based on the corresponding imaging device parameters,

wherein the prediction of the depth map comprises predicting also an uncertainty map relating to the depth map.

2. The method according to claim 1 , wherein the data of the training set has higher resolution that the at least two 2D satellite images.

3. The method according to claim 1 , wherein the multi-view 3D geocoded ground truth data is real world data.

4. The method according to claim 1 , wherein the multi-view 3D geocoded ground truth data is extracted from a geocoded 3D surface model for training purposes, said geocoded 3D surface model being provided based on real world data.

5. The method according to claim 1 , wherein the multi-view 3D geocoded ground truth data comprises images and/or LIDAR measurement data and/or radar measurement data and/or sonar measurement data.

6. The method according to claim 1 , wherein the overlapping 2D satellite images are bundle adjusted.

7. The method according to claim 1 , wherein the training set relates to a built up environment.

8. The method according to claim 1 , wherein the multi-view 3D geocoded ground truth data relates to a plurality of geographical areas and comprises a plurality of reference images and/or measurement data capturing each geographical area from different angles.

9. The method according to claim 1 , wherein the ground truth data is geocoded.

10. The method according to claim 1 , further comprising a step of training ( 99 ) the at least one MLN.

11. A non-transitory computer-readable medium for 3D reconstruction from satellite imagery using deep learning, comprising instructions which, when executed by at least one processor cause the at least one processor to carry out the method according to claim 1 .

12. A method for 3D reconstruction from satellite imagery using deep learning, said method comprising the steps of:

providing ( 101 ) at least two overlapping 2D satellite images,

providing ( 102 ) imaging device parameters for the at least two overlapping 2D satellite images,

providing ( 103 ) at least one trained Machine Learning Network, MLN, able to predict depth maps, said trained MLN being trained on a training set comprising multi-view 3D geocoded ground truth data, and

predicting ( 104 ) a depth map of the at provided at least two 2D satellite images using the trained at least one MLN and based on the corresponding imaging device parameters,

wherein the MLN comprises a plurality of MLNs each predicting a depth map and wherein an updated depth map is formed based on the predicted depth maps and when available based on the uncertainty map relating to the respective depth map.

13. The method according to claim 12 , wherein one or more of:

the data of the training set has higher resolution that the at least two 2D satellite images,

the multi-view 3D geocoded ground truth data is real world data,

the multi-view 3D geocoded ground truth data is extracted from a geocoded 3D surface model for training purposes, said geocoded 3D surface model being provided based on real world data,

the multi-view 3D geocoded ground truth data comprises images and/or LIDAR measurement data and/or radar measurement data and/or sonar measurement data,

the overlapping 2D satellite images are bundle adjusted,

the training set relates to a built up environment,

the multi-view 3D geocoded ground truth data relates to a plurality of geographical areas and comprises a plurality of reference images and/or measurement data capturing each geographical area from different angles, or

the ground truth data is geocoded.

14. A method for 3D reconstruction from satellite imagery using deep learning, said method comprising the steps of:

providing ( 101 ) at least two overlapping 2D satellite images,

providing ( 102 ) imaging device parameters for the at least two overlapping 2D satellite images,

providing ( 103 ) at least one trained Machine Learning Network, MLN, able to predict depth maps, said trained MLN being trained on a training set comprising multi-view 3D geocoded ground truth data, and

predicting ( 104 ) a depth map of the at provided at least two 2D satellite images using the trained at least one MLN and based on the corresponding imaging device parameters,

wherein the prediction of the depth map comprises associating the data of the depth map to geocoded coordinate data.

15. The method according to claim 14 , wherein one or more of:

the data of the training set has higher resolution that the at least two 2D satellite images,

the multi-view 3D geocoded ground truth data is real world data,

the multi-view 3D geocoded ground truth data is extracted from a geocoded 3D surface model for training purposes, said geocoded 3D surface model being provided based on real world data,

the multi-view 3D geocoded ground truth data comprises images and/or LIDAR measurement data and/or radar measurement data and/or sonar measurement data,

the overlapping 2D satellite images are bundle adjusted,

the training set relates to a built up environment,

the multi-view 3D geocoded ground truth data relates to a plurality of geographical areas and comprises a plurality of reference images and/or measurement data capturing each geographical area from different angles, or

the ground truth data is geocoded.

16. A method for 3D reconstruction from satellite imagery using deep learning, said method comprising the steps of:

providing ( 101 ) at least two overlapping 2D satellite images,

providing ( 102 ) imaging device parameters for the at least two overlapping 2D satellite images,

providing ( 103 ) at least one trained Machine Learning Network, MLN, able to predict depth maps, said trained MLN being trained on a training set comprising multi-view 3D geocoded ground truth data,

predicting ( 104 ) a depth map of the at provided at least two 2D satellite images using the trained at least one MLN and based on the corresponding imaging device parameters, and

generating ( 105 ) a geocoded 3D surface model based on the predicted depth map.

17. The method according to claim 16 , wherein one or more of:

the data of the training set has higher resolution that the at least two 2D satellite images,

the multi-view 3D geocoded ground truth data is real world data,

the multi-view 3D geocoded ground truth data is extracted from a geocoded 3D surface model for training purposes, said geocoded 3D surface model being provided based on real world data,

the multi-view 3D geocoded ground truth data comprises images and/or LIDAR measurement data and/or radar measurement data and/or sonar measurement data,

the overlapping 2D satellite images are bundle adjusted,

the training set relates to a built up environment,

the multi-view 3D geocoded ground truth data relates to a plurality of geographical areas and comprises a plurality of reference images and/or measurement data capturing each geographical area from different angles, or

the ground truth data is geocoded.

18. A method for 3D reconstruction from satellite imagery using deep learning, said method comprising the steps of:

providing ( 101 ) at least two overlapping 2D satellite images,

providing ( 102 ) imaging device parameters for the at least two overlapping 2D satellite images,

providing ( 103 ) at least one trained Machine Learning Network, MLN, able to predict depth maps, said trained MLN being trained on a training set comprising multi-view 3D geocoded ground truth data, and

predicting ( 104 ) a depth map of the at provided at least two 2D satellite images using the trained at least one MLN and based on the corresponding imaging device parameters,

wherein the trained MLN is arranged to predict depth maps for built-up environments.

19. The method according to claim 18 , wherein one or more of:

the data of the training set has higher resolution that the at least two 2D satellite images,

the multi-view 3D geocoded ground truth data is real world data,

the multi-view 3D geocoded ground truth data is extracted from a geocoded 3D surface model for training purposes, said geocoded 3D surface model being provided based on real world data,

the multi-view 3D geocoded ground truth data comprises images and/or LIDAR measurement data and/or radar measurement data and/or sonar measurement data,

the overlapping 2D satellite images are bundle adjusted,

the training set relates to a built up environment,

the multi-view 3D geocoded ground truth data relates to a plurality of geographical areas and comprises a plurality of reference images and/or measurement data capturing each geographical area from different angles, or

the ground truth data is geocoded.

20. A method ( 300 ) for training a Machine Learning Network, MLN, for 3D reconstruction from satellite imagery, said method comprising the steps of:

providing ( 301 ) at least one trained Machine Learning Network, MLN, able to predict depth maps, and

training ( 302 ) said MLN on a training set comprising at least one textured or non-textured 3D geocoded 3D surface model for training purposes, wherein multi-view 3D geocoded ground truth data is rendered from said textured or non-textured geocoded 3D surface model for training purposes.

21. A non-transitory computer-readable medium for training a Machine Learning Network, MLN, for 3D reconstruction from satellite imagery, comprising instructions which, when executed by at least one processor cause the at least one processor to carry out the method according to claim 20 .