Information Technology Reference
In-Depth Information
70
60
50
40
30
20
10
0
0
100
200
300
400
500
600
700
Images added
Fig. 3. The time required to optimize a large scene by incrementally adding images,
showing the approximately quadratic relation between scene size and calculation speed
Table 1. Evaluation of the influence of the partitioning size S ,for O = 10. The total
time to optimize the scene using the methods from literature [12] was 2829 seconds.
All timing results were achieved on an Intel Core 2 Duo T9300 CPU. The total scene
size is about 25 x 20 x 10 units. See the text for details.
S
time subset (s)
time total (s)
μ
σ
μ opt
σ opt
20
70
667
0.02315
0.01302
0.01248
0.00921
30
147
685
0.01937
0.01008
0.00863
0.00750
40
221
745
0.03692
0.02054
0.01032
0.00978
50
314
836
0.01655
0.01027
0.01520
0.01112
60
408
1001
0.00841
0.00527
0.00232
0.00241
70
492
1341
0.00742
0.00825
0.00320
0.00419
time required when the method runs sequential, in case only 1 CPU core would
be available. Note that for all values of S the total running time is lower than
when we would not split the scene. This confirms our observations from figure 3.
The next four columns give accuracy results. The first μ and σ are the mean and
standard deviation of all the differences between the 3D positions of all points
in the reconstructed scene using our method and the method from literature.
The second μ opt and σ opt are the results after running a bundle adjustment on
the recombined scenes. This fixes deviations introduced by the splitting, but of
course at the cost of some computation time. The presented numbers only make
sense in relation to the total scene size, which is about 25 x 20 x 10 units. The
reported errors are thus a factor 10 4 smaller than the scene size, meaning that
the scene reconstructed with our method will be visually identical to the scene
reconstructed using state of the art methods.
Search WWH ::




Custom Search