Implementation of a Self-Consistent Stereo
Description: Implementation of a Self-Consistent Stereo Processing Chain for 3D Stereo Reconstruction of the Lunar Surface E. Tasdelen1, H. Unbekannt1, M. Yildirim1, K. Willner1 and J. Oberst1,2 1 Department of Geodesy and Geoinformation Science,
Related Topics
Download Presentation
"Implementation of a Self-Consistent Stereo" is the property of its rightful owner. Permission is granted to download and print the materials on this website for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Presentation Transcript
slide1. Implementation of a Self-Consistent Stereo Processing Chain for 3D Stereo Reconstruction of the Lunar Surface E. Tasdelen1, H. Unbekannt1, M. Yildirim1, K. Willner1 and J. Oberst1,2 1 Department of Geodesy and Geoinformation Science, Technical University of Berlin
2 German Aerospace Center (DLR)<br>
slide2. Motivation The department for Planetary Geodesy at TU Berlin is developing routines for photogrammetric processing of planetary image data to derive 3D representations of planetary surfaces.
Aim: An independent generic 3D reconstruction pipeline
Integrated Software for Imagers and Spectrometers (ISIS) developed by USGS Flagstaff, was chosen as a prime processing platform and tool kit.<br>
slide3. Matching Software Overview of the software Supports multithreading
Improved performance
Memory management for large images
Image formats
Vicar, ISIS cube, TIFF<br>
slide4. Matching Algorithms Reference Image Search Image where is covariance
are variances Area-based Matching (ABM) source: Rodehorst, 2004 Normalized Cross-Correlation (NCC)<br>
slide5. Matching Algorithms Reference Patch Compared Patches Least-Squares Matching (LSM): source: Bethmann et al., 2010 Functional Model: f(x,y) + e(x,y) = g(x’,y’) Transformation Model: x = a0 + a1x’ + a2y’
y = b0 + b1x’ + b2y’ a0 + a1x’ + a2y’ 1 + c1x’ + c2y’ x = b0 + b1x’ + b2y’ 1 + c1x’ + c2y’ y = Projective transformation<br>
slide6. Matching Types Type1: Matching images without pre-processing
Same search space for each pixel Type2: Coarse-to-fine hierarchical matching
Results from the pyramids override the search space boundaries<br>
slide7. Matching Types Type3: Grid-based matching
Grid-based projective transformation<br>
slide8. Blunder Detection The main reasons of blunders
occlusions, depth discontinuities, repetitive patterns, inadequate texture, etc.
Filters
Epipolar Check: With the help of epipolar geometrical relation, all the matched points are controlled and the distances of the points to the corresponding epipolar lines are calculated. Points exceeding a set threshold distance to the epipolar line are discarded. Epipolar Relation Epipolar Error Check<br>
slide9. Blunder Detection Overlapping Area Check: divide the reference image into regular sized grids and check if there are adequate numbers of tie-points within each grid. (a-b) left and right pair of stereo images, (c) actual overlapping area visualized on the left image, (d-f) grids with different sizes on the left image (300, 200 and 100 from d to f, respectively)<br>
slide10. LRO NAC Images for Copernicus Crater Resulting Disparity Map 49750593
correspondences 1km -500PX 150PX<br>
slide11. 3D Point Calculation Forward Ray Intersection
Computation of spatial object coordinates X from measured image points x and x’ as well as the camera matrices P and P’. source: Rodehorst, 2004<br>
slide12. Blunder Detection Filters on 3D point data
Octree Filter: uses octree data structure created from 3D point cloud data. Nodes with low density, containing only few points, are considered as noise source: Wang, 2012<br>
slide13. Blunder Detection Filters on 3D point data
Delaunay Triangles: Each point is connected by lines to its closest neighbors, in such a way The points which contributes triangles with edge length exceeding a threshold indicates the possible outliers.<br>
slide14. DTM Interpolation 3D point coordinates are first map-projected to a grid based images
Colliding points are interpolated
IDW, nearest neighbor, mean or median
A customized search radius can be applied to define the pixel value.<br>
slide15. Main Challenges:
Rendering capabilities of graphics hardware
Limited to several millions of primitives per second
Geometry throughput effects the performance
Tremendous size of data does not fit into memory
Ex: 15km x 15km area with 1.5m res. > 5 GB of data, simply cannot be placed into memory at once Visualization Tool [1] source: Wang, 2012<br>
slide16. Visualization Tool Level Of Detail (LOD) Algorithm
Decreasing the complexity of the object with the increasing distance to the viewer source: Bekiaris, 2009<br>
slide17. Visualization Tool Surface Representation Simplification Level Of Detail (LOD) Algorithm
Based on Quad Trees Each child chunk represent a more detailed version of one of its parents quarters Each segment is called as a chunk source: Ulrich, 2002<br>
slide18. Visualization Tool LOD 1 Viewer LOD 2 LOD 0 Representation Rendering wrt. viewing direction<br>
slide19. Landing Module 72.195 km<br>
slide20. Landing Module ~1000m @Landing Module The position of Apollo 17 landing module<br>
slide21. Landing Module ~1000m @Landing Module The position of Apollo 17 landing module<br>
slide22. A look towards south from the position of Apollo 17 landing module<br>
slide23. A look towards north from the position of Apollo 17 landing module<br>
slide24. Matching Algorithms NCC: correlation coefficient correlation position 0.0 1.0 0.5 Maximum
Correlation Problems:
NCC is not defined for homogeneous image areas(variance is zero)!
NCC is not invariant to geometrical distortions
Pixel accuracy! Threshold = 0.8<br>
slide25. Visualization Tool<br>
2 German Aerospace Center (DLR)<br>
slide2. Motivation The department for Planetary Geodesy at TU Berlin is developing routines for photogrammetric processing of planetary image data to derive 3D representations of planetary surfaces.
Aim: An independent generic 3D reconstruction pipeline
Integrated Software for Imagers and Spectrometers (ISIS) developed by USGS Flagstaff, was chosen as a prime processing platform and tool kit.<br>
slide3. Matching Software Overview of the software Supports multithreading
Improved performance
Memory management for large images
Image formats
Vicar, ISIS cube, TIFF<br>
slide4. Matching Algorithms Reference Image Search Image where is covariance
are variances Area-based Matching (ABM) source: Rodehorst, 2004 Normalized Cross-Correlation (NCC)<br>
slide5. Matching Algorithms Reference Patch Compared Patches Least-Squares Matching (LSM): source: Bethmann et al., 2010 Functional Model: f(x,y) + e(x,y) = g(x’,y’) Transformation Model: x = a0 + a1x’ + a2y’
y = b0 + b1x’ + b2y’ a0 + a1x’ + a2y’ 1 + c1x’ + c2y’ x = b0 + b1x’ + b2y’ 1 + c1x’ + c2y’ y = Projective transformation<br>
slide6. Matching Types Type1: Matching images without pre-processing
Same search space for each pixel Type2: Coarse-to-fine hierarchical matching
Results from the pyramids override the search space boundaries<br>
slide7. Matching Types Type3: Grid-based matching
Grid-based projective transformation<br>
slide8. Blunder Detection The main reasons of blunders
occlusions, depth discontinuities, repetitive patterns, inadequate texture, etc.
Filters
Epipolar Check: With the help of epipolar geometrical relation, all the matched points are controlled and the distances of the points to the corresponding epipolar lines are calculated. Points exceeding a set threshold distance to the epipolar line are discarded. Epipolar Relation Epipolar Error Check<br>
slide9. Blunder Detection Overlapping Area Check: divide the reference image into regular sized grids and check if there are adequate numbers of tie-points within each grid. (a-b) left and right pair of stereo images, (c) actual overlapping area visualized on the left image, (d-f) grids with different sizes on the left image (300, 200 and 100 from d to f, respectively)<br>
slide10. LRO NAC Images for Copernicus Crater Resulting Disparity Map 49750593
correspondences 1km -500PX 150PX<br>
slide11. 3D Point Calculation Forward Ray Intersection
Computation of spatial object coordinates X from measured image points x and x’ as well as the camera matrices P and P’. source: Rodehorst, 2004<br>
slide12. Blunder Detection Filters on 3D point data
Octree Filter: uses octree data structure created from 3D point cloud data. Nodes with low density, containing only few points, are considered as noise source: Wang, 2012<br>
slide13. Blunder Detection Filters on 3D point data
Delaunay Triangles: Each point is connected by lines to its closest neighbors, in such a way The points which contributes triangles with edge length exceeding a threshold indicates the possible outliers.<br>
slide14. DTM Interpolation 3D point coordinates are first map-projected to a grid based images
Colliding points are interpolated
IDW, nearest neighbor, mean or median
A customized search radius can be applied to define the pixel value.<br>
slide15. Main Challenges:
Rendering capabilities of graphics hardware
Limited to several millions of primitives per second
Geometry throughput effects the performance
Tremendous size of data does not fit into memory
Ex: 15km x 15km area with 1.5m res. > 5 GB of data, simply cannot be placed into memory at once Visualization Tool [1] source: Wang, 2012<br>
slide16. Visualization Tool Level Of Detail (LOD) Algorithm
Decreasing the complexity of the object with the increasing distance to the viewer source: Bekiaris, 2009<br>
slide17. Visualization Tool Surface Representation Simplification Level Of Detail (LOD) Algorithm
Based on Quad Trees Each child chunk represent a more detailed version of one of its parents quarters Each segment is called as a chunk source: Ulrich, 2002<br>
slide18. Visualization Tool LOD 1 Viewer LOD 2 LOD 0 Representation Rendering wrt. viewing direction<br>
slide19. Landing Module 72.195 km<br>
slide20. Landing Module ~1000m @Landing Module The position of Apollo 17 landing module<br>
slide21. Landing Module ~1000m @Landing Module The position of Apollo 17 landing module<br>
slide22. A look towards south from the position of Apollo 17 landing module<br>
slide23. A look towards north from the position of Apollo 17 landing module<br>
slide24. Matching Algorithms NCC: correlation coefficient correlation position 0.0 1.0 0.5 Maximum
Correlation Problems:
NCC is not defined for homogeneous image areas(variance is zero)!
NCC is not invariant to geometrical distortions
Pixel accuracy! Threshold = 0.8<br>
slide25. Visualization Tool<br>