3D-matic and EPCC jointly developed Grid strategies for the
stereo-correlator, the most time-intensive part of the 3D image
processing. The task lends itself well to distributed processing on an
image-pair basis.
The stereo-correlator works in two phases:
- Matching: pairs of monchrome images are compared to extract range information.
- Integration: the range information obtained from a number of camera
pairs is integrated into a voxel structure, combined with colour images
obtained by a number of colour cameras, to yield a triangulated 3D
model.
The frame sequencing generates new matching and model building work continuously.
The most effective way to handle such a stream of work in an environment where
processing resources are constantly entering and being removed from the available
pool is to provide a mechanism for dynamically spawning matching and building
tasks remotely on appropriate process servers. Furthermore, some spawned tasks
may need to create and maintain data connections to others on other hosts
(e.g. matcher servers dispatch the range maps for a frame to the server which
has been selected to perform the model build for that frame).
The dynamic task creation and remote pipe migration required by PGPGrid motivates
the development of a Java interface loosely modelled on the primitives of Milners's
pi-calculus. The interface, JPie, is intended for use as a substratum for this and
other GRID-based parallel computing applications. It allows the dynamic creation of
both remote processes and communications channels between them. It also allows
dynamic reconfiguration of the network of channels. JPie tries to achieve this
with the minimum number of primitives, integrating these into the existing
Java class framework.
Download the JPie inteface Java code (Zip archive, 256 Kbytes).
Please mail us your comments.
|
