The City Scanning Project :
69
The City Scanning Project : Validation and parallel algorithms Olivier Koch, Visitor from March to August 2002 Advisor : Seth Teller
description
The City Scanning Project :. Validation and parallel algorithms. Olivier Koch, Visitor from March to August 2002. Advisor : Seth Teller. Contents. 1/ Project overview 2/ code validation 3/ parallel algorithms 4/ a few ideas. - PowerPoint PPT Presentation
Transcript of The City Scanning Project :
The City Scanning Project :
Validation and parallel algorithms
Olivier Koch,Visitor from March to August 2002
Advisor : Seth Teller
Contents
1/ Project overview
2/ code validation
3/ parallel algorithms
4/ a few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Reconstruct a realistic 3D model of an urban environment from calibrated real-world images
Argus
Technology Square
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• Nodes
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• Edge detection
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
edge
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
edge
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
edge
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
Horizontal plane
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
Horizontal plane
Horizontal line
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
Horizontal plane
Horizontal line
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
Horizontal plane
horizontal line
horizontal line
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Camera plane
Horizontal plane
Horizontal line
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Camera centerGround plane
Horizontal line
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Ground plane
Horizontal line
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• assume edges are horizontal• determine corresponding orientation in 3D world
Ground plane
Horizontal line
X
Y
Z
= azimuth of 3D edge
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• Edge detected…
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• Edge detected… now have an azimuth !
• each azimuth is color-coded.
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• histogram edges
0 360
azimuth
weight
Façade orientation wrt due east
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• histogram horizontal edges
0 360
azimuth
weight
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• histogram horizontal edges
0 360
azimuth
weight
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• keep orientations that are supported by at least 3 nodes
0 360
azimuth
weight
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• determine corresponding edges
0 360
azimuth
weight
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
From : Extracting Textured Vertical Facades from Controlled Close-Range Imagery, S. Teller, S. Coorg in Proc. CVPR, pp. 625-632, June 1999.
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
grid cell
correlation
projected edge
computed correlation value
step
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
sweep step
grid cell
correlation
step
peak
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
sweep step
grid cell
correlation
step
peak
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
grid cell
correlation
step
peak
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
11
1 1
11
2
2
2
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Project overview
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Contents
1/ Project overview
2/ code validation
3/ parallel algorithms
4/ a few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Code validation
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• edge detection improvement
automatic edges manual edges
Code validation
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• edge detection improvement
automatic edges
manual edges
Code validation
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• polygon filtering : remove polygons without support.
without filter with filter
Code validation
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• debugging
warnings
bugs
Rational Purify
CVD
log files, class constructors, …
Contents
1/ Project overview
2/ code validation
3/ parallel algorithms
4/ a few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• why parallelize ?not processed
being processed
processed
IRIX machine
grid cells
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• why parallelize ?not processed
being processed
processed
IRIX machine
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• why parallelize ?not processed
being processed
processed
IRIX machine
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• why parallelize ?not processed
being processed
processed
> 7 hours on IRIX single processor=
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• first method
Machine 1Machine 2
Machine 3
Machine 5
Machine 4
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• first method• CPU speed varies across machines• time to process a cell is not constant
difficult optimization problem…
turpentineraytracemosaicacetonehuepanoramaorange
2285 sec.2678244321312711234217551939
Computing time : 45 min.Standard dev : 3.4 %
Static allocation
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
Machine 1Machine 2
Machine 3
Machine 5
Machine 4
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
Machine 1Machine 2
Machine 3
Machine 5
Machine 4
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
Machine 1Machine 2
Machine 3
Machine 5
Machine 4
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
Machine 1Machine 2
Machine 3
Machine 5
Machine 4
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
Machine 1Machine 2
Machine 3
Machine 5
Machine 4
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
Machine 1Machine 2
Machine 3
Machine 5
Machine 4
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
• master-worker model (MPI)
master
worker #1 worker #2
worker #3
worker #4worker #5
masterworker“I am free”
masterworker
“processcell #xx”
masterworker
“cell #xx hasbeen processed”
Communication scheme
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• second method
panorama
turpentine
acetone
mosaic
trace
orangehueray
1246.31246.31248.51248.61250.01240.61250.31247.01243.01233.81249.11239.11250.51238.31255.61278.91276.01275.91277.21268.81230.4
• Computation time 21mn
(sec.)
• Idle time 6mn
• Communication time3mn
Standard deviation : 2.6 %
Parallel algorithms
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Method 1(rectangles)
Method 2(queue)
Time Speed-up
• method 2 is faster and more flexible.• last step of reconstruction still needs to be parallelized
45 min. 10.0
30 min. 15.2
(takes ~10 minutes on a single processor IRIX)
Contents
1/ Project overview
2/ code validation
3/ parallel algorithms
4/ a few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
A few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• high-frequency edge filtering
A few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• high-frequency edge filtering
A few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• high-frequency edge filtering
boolean image high-frequency detection
mask
- on courtesy of Alexis Charignon
A few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• high-frequency edge filtering
A few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• camera position refinement
Argus
Argus position
Argus position
A few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• camera position refinement
Argus position
Argus position
• orange robots keep track of relative Argus motion.
A few ideas
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
• using alternative sensors
IR imagery
Infrared Services Inc.
Conclusion
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Infrared Services Inc.
Conclusion
• reconstruction code validated on Tech Square
• most CPU-intensive program parallelized
• speed-up = 15 on IRIX cluster
Next steps
• refining camera positions
• porting the code to Linux
Acknowledgements
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Infrared Services Inc.
• Seth Teller
• Manish Jethwa, Neel Master
• Alexis, Yann, Stefane, Adnan
• MIT Graphics Group
The End
The City Scanning Project – validation and parallel algorithms – July 2002 – [email protected]
Infrared Services Inc.
Thanks…
Questions ?
