Graphics Reference
In-Depth Information
6.5.2 Acceleration by Elevated Granularity
By elevating the processing granularity from pixels to tiles, the algorithm execution
speed can be drastically accelerated. In general, the computational complexity of the
processing is inversely proportional to the granularity of the tessellation. If the tile
size is chosen wisely, the speed can be significantly increased without noticeable
visual quality impact. Our system uses three optimization schemes based on these
speed versus quality trade-offs.
Tiled Undistortion Standard lens distortion is generally corrected on a pixel-
basis level, but can be approximated by applying an equivalent geometrical undistor-
tion to small image tiles using a resolution factor 0
1. Since a GPU pipeline
exists out of a geometry and pixel processing stage, the lens correction can hence
be ported from the pixel to the geometry stage. The pixel processing stage becomes
clear to perform the consecutive segmentation processing in a single pipeline pass,
which significantly leverages the GPU utilization.
Tiled Tallying in Reduced Bins For the movement analysis, the multiresolution
capabilities of the GPU are used to tally tiles instead of pixels in the histogram bins.
This sampling resolution is expressed by a factor 0
tu
s
1, where
ˁ s is proportional
to the granularity of the tiles.
Evidently, it is of no use to have more histogram bins than the number of planes
in the sweep. However, the essential part is deriving the parameters
to adjust
the dynamic range of the depth scan. As depicted in Fig. 6.11 c, d, we are able to
approximate the histogram by reducing the number of bins, without a large impact
on the Gaussian parameters. Therefore, the number of bins are defined proportional
to the number of planes M , with factor 0
μ
and
˃
b
1. Heavily reducing the number
of bins (see Fig. 6.11 d) causes the center
to become less accurate, as it is shifted
toward the center of the effective scan range. An optimal trade-off point can therefore
be defined, since the accuracy loss will cause the responsiveness of the system to
decrease.
Tiled Splatting Identical to the tiled undistortion, the depth map can be tes-
sellated with a factor 0
μ
1 to form a mesh for splatting tiles instead of
pixels. This technique can significantly accelerate the confident camera recoloring,
by interpolating angles between the tile corners.
ts
6.6 Results
We demonstrate our previously discussed methods using a prototype setup. Our pro-
totype setup is built with N
6 auto-synchronized Point GrayResearchGrasshopper
cameras mounted on an aluminum frame, which closely surrounds the screen (see
Fig. 6.1 , top right). The presented camera setup avoids large occlusions, and has the
potential to generate high quality views since no image extrapolation is necessary.
We have used the Multicamera Self-Calibration toolbox [ 27 ] to calibrate the camera
=
Search WWH ::




Custom Search