A "Region-Growing" Algorithm for Matching of Terrain Images.
read more
Citations
Development and Status of Image Matching in Photogrammetry
Automatic Digital Surface Model (DSM) generation from linear array images
Scene Reconstruction and Visualization From Community Photo Collections Recent progress is described in digitizing and visualizing the world from data captured by people taking photos and uploading them to the web.
Algorithm Theoretical Basis
Automatic dsm generation from linear array imagery data
Related Papers (5)
Using synthetic images to register real images with surface models
PMF: A Stereo Correspondence Algorithm Using a Disparity Gradient Limit:
Frequently Asked Questions (16)
Q2. How long does it take to match a patch?
then, matching one pair of patches on a Transputer will typically take between 0.05s and 0.2s, depending on the parameters used, and the image characteristics.
Q3. What is the error in the DEM derived from the underflight photography?
The DEM derived from the underflight photography is estimated to have an RMS height error of about l-2m; in addition, much of the remaining error is due to the sensor model used.
Q4. How many constant expressions can be added to the old?
by moving constant expressions out of the loop & so forth, the new values of x_from, y_from can be calculated from the old with just one addition each.
Q5. How many processors can parallelise the algorithm?
The authors have already parallelised their algorithm on a network of SUN workstations, and speedup is, as predicted, linear up to 15 processors (which was all the authors could get their hands on!).
Q6. What is the main item not specified by this pseudo-code?
Selecting which match should be used to grow from ...The remaining major item not specified by this pseudo-code is which match should be selected from list_to_be_grown_from on each iteration.
Q7. How many control points are used to calculate the y disparity?
the SPOT satellite's attitude varies slightly & not very predictably during its orbit, so that it is very difficult to calculate the y disparity to better than a pixel or so, unless hundred's of control points are used.
Q8. How many interations does it take to get a good theoretical estimate?
The number of interations required depends heavily upon the input data (image & initial parameter values), so it is very difficult to get a good theoretical estimate.
Q9. Who helped us with the Transputer-related issues?
The authors would also like to thank their colleagues for help and useful discussions; most notably, Kevin Collins (of RSRE) helped us with Transputer-related issues, while Tim Day and JP.
Q10. What is the reason for linearizing in the way Gruen does?
"One of the reasons for linearizing in the way Gruen does is that the equations are then in the form of a "Gauss-Markov estimation model".
Q11. What is the remaining error due to the algorithm?
The remaining error is that due to the matching algorithm; this error appears to be significantly less than 0.5 pixels (RMS), possibly as low as 0.1 pixels (RMS).
Q12. How many pixels are in the original?
This figure is based on: each point requiring about 0.2 CrV-seconds to match (on average); matching all the points on a square grid consisting of every 6th pixel horizontally or vertically (ie. 1 in 36 pixels of the original); 6000 x 6000 images.
Q13. What is the way to divide the distortions into two classes?
It is convenient to divide the distortions into two classes: radiometric (which affect the measured grey-level at any corresponding point), and geometric (which affect the positions of corresponding points).
Q14. how long does it take to produce a high-quality, dense disparity map?
Summary — speedUsing 30 T800 Transputers, and the algorithm above, it should be possible to produce a high-quality, dense disparity map from a pair of SPOT images in about 2 hours.
Q15. What can be used to help in the initial "Seeding" of the algorithm?
(Though such knowledge can be used to get more accurate matches, and to assist in the initial "seeding" of the algorithm with approximate matches.
Q16. How many approximate correspondences do the authors need to seed their algorithm?
The authors have used only small numbers of approximate correspondences to seed their algorithm (typically 3 or 4); this has worked well so far, but the authors will probably use more when the authors automate this stage properly.