Mars Image Processing

Correlation Mode Dependence

Introduction:

The correlation algorithm has several lower level correlation modes available to the user. The socalled amoeba algorithm performshe a correlation matrix elvaluation in a siz dimensional space that allows for dispalcement as well as image resizing. The amoeba2 algorithm only considers a two dimensional displacement problem, which is quite appropriate for our case considered here for the stereo image correlation. Various other modes are also alvailable. With the availability of a correlation quality measurement we can no compare the output of various algorithms and evaluate the number of pixels returned by each algorithm measured against the amoeba algorithm. The results for image set 1 and image set 2 are tabulated below.

Image Set 2:

mode time All
Pixels		 Good
Pixels		 Bad
Pixels		 Pixels
in
Neither		 Pixels
in
Reference
Only		 Pixels
in
Mode
Only		 Line
Pixels
0 pixel
Deviation		 Line
Pixels
1 pixel
Deviation		 Line Pixel
2 pixel
Deviation		 Line
Pixel
>2 pixel
Deviation		 Sample
Pixels
0 pixel
Deviation		 Sample
Pixels
1 pixel
Deviation		 Sample
Pixels
2 pixel
Deviation		 Sample
Pixels
>2 pixel
Deviation
amoeba 149.49 307200 227774 79426 0 0 0 227774 0 0 0 227774 0 0 0
amoeba2 22.87 307200 217400 89800 77678 12122 1748 214422 1197 31 2 211973 3567 74 38
lin 55.23 307200 200409 106791 77927 28864 1499 196347 2501 41 21 195653 3188 43 26
lin_am 167.461 307200 227048 80152 78700 1452 726 226185 131 4 2 226184 74 25 39
lin_am2 68.34 307200 217369 89831 77778 12053 1648 214456 1235 19 11 212004 3598 78 41

Image Set 1:

mode time All
Pixels		 Good
Pixels		 Bad
Pixels		 Pixels
in
Neither		 Pixels
in
Reference
Only		 Pixels
in
Mode
Only		 Line
Pixels
0 pixel
Deviation		 Line
Pixels
1 pixel
Deviation		 Line Pixel
2 pixel
Deviation		 Line
Pixel
>2 pixel
Deviation		 Sample
Pixels
0 pixel
Deviation		 Sample
Pixels
1 pixel
Deviation		 Sample
Pixels
2 pixel
Deviation		 Sample
Pixels
>2 pixel
Deviation
amoeba 145.347 307200 136314 170886 0 0 0 136314 0 0 0 136314 0 0 0
amoeba2 22.523 307200 134747 172453 161108 11345 9778 122841 1926 125 77 122704 2059 113 93
lin 47.004 307200 134419 172781 156905 15876 13981 117726 2335 120 257 117804 2211 105 318
lin_am 156.165 307200 140828 166372 159250 7122 11636 128188 819 27 158 128692 248 24 228
lin_am2 68.104 307200 139992 167208 156703 10505 14183 122967 2516 127 199 123243 2136 102 338

Conclusion:

The use of the amoeba2 algorithm deduces the required CPU time dramatically from approximately 150 seconds to approximately 23 seconds (on 25 CPUs) with a loss of pixels of 5-10%.

Acknowledgements:

This work was sponsored by the TMOD technology program under the Beowulf Application and Networking Environment (BANE) task.The original VICAR based software is maintained in the Multi-mission Image Processing Laboratory (MIPL). The work was performed in a collaboration between Gerhard Klimeck, Myche McAuley, Tom Cwik, Bob Deen, and Eric DeJong

Mars Image Processing

Mars Home Mosaics Stereo Correlation Parallel Scaling 8,16 CPUs: Load Imbalance 39 CPUs: Staging Problems Quality Measurement Mode Dependence