MIME-Version: 1.0
Server: CERN/3.0
Date: Monday, 25-Nov-96 00:16:39 GMT
Content-Type: text/html
Content-Length: 4952
Last-Modified: Tuesday, 19-Sep-95 22:26:57 GMT
Suggested Robotlab M.Eng Projects
Suggested Robotlab M.Eng Projects
Robotics
- Wireless Robotics-Get Rosemary to communicate with the Sun workstations using the wireless radio modem, with wireless video (using a cheap box which
sends an analog video signal over radio (like TV)). Hook this up to
the ATM video source and build an application of realtime vision
tracking/etc. with rosemary as the host. (Matt)
- Visual Robot Navigation-An extension to Tom's project:
Mapmaking with visual cues. Can use ideas from (1), but the robot
walks around, and, using depth and position information produces a map
of the world, uses it to navigate. (Matt) [Tom's project: a set of navigational becons were created for the robots]
Video Processing and Machine Vision
- Automatic Storyboarding - use the dissolve detector code to
automate generation of movie storyboards, possible outputing html
pages that show the story board. (Justin and Ramin)
- Parallelization of the Mpeg encoder Write a 'real-time'
(or at least very fast) Mpeg encoder using the ATM cluster/SP2 and ATM
video source. Jon Swartz and Brian Smith were interested in this. (Matt)
- Port Split-C to Windows NT. This could prove to be
non-trivial, as many aspects of the SMP design rely on a UNIX process
environment, in particular ELF, and shared memory. Use of NT features
such as threads won't work. (Matt and Justin V.)
- Parallel RIVL. Providing Split-C/Active Messages support
for RIVL. We'd like RIVL to be the general framework for going image
computation, display, applications building, etc. (It should help
productivity a lot if we have an easy way for people to experiment and
visualize vision algorithms, instead of writing crufty C programs.)
This project would provide fine-grain data parallel processing support
for the RIVL architecture, basically by (a) porting various RIVL
image-processing primitives to Split-C (for automatic parallelization)
and (b) allowing users to write new RIVL primitives in Split-C for
linking into the system. The easiest way to do all of this, I think:
Port RIVL itself to Split-C (should be easy) and make existing RIVL
behaviour a subset of the "Split-RIVL" (that is, local processing on
one machine). (Matt)
- Effects of Compression What (empirical) effects do various compression algorithms have on common vision algorithms. (Justin and Ramin)
- Edge Clustering Develop algorithm for clustering of
edges that represent a single object. (Justin and Ramin)
- Adobe Photoshop plug-ins(Nawaaz)
Algorithms
- OptimizationCreate an environment
for the automatic optimization of highly iterative image
transformation algorithms using the m4 macro language and perl
scripts. Write m4 macros to automatically produce C-code according to
some set of parameters (like how many times to unroll a loop) and the
write a perl script to compile the set of programs and determine which
is the fastest by timing the transformation on some set of images.
This would greatly speed the production of optimized code for a given
machine. I have examples started of this idea for the rank transform.
- Search-based compilation
Modelling compilation as a search problem on a system with constraints :
The system will consist of
- a set of processors
- a set of memories
- an interconnect network
- a set of constraints
- number of processors
- size of memory
- which processor can access which memories in a cycle
- ordering imposed by the operations to be performed.
- etc.
The code to be parellised will be tight loops and hence small.
The idea is to search the space of complied code to find the optimal one.
This is most probably NP complete, but speed is not the essence here.(Nawaaz)
- Performance Simulations The simulation is with respect
to the algorithms we are interested in. The many aspects that can be
simulated are : a) Cache -- implement and study cache accesses on
various algorithms. Here we assume a single processor. b) Shared
Memory Multiprocesssors -- simulate and study the effect of bus
bandwidth, memory architecture , cache coherence on the system (with
repect to alogrithms like correlation etc) (Nawaaz)
- DSPs Implementation / Compilation / Code generation for
DSP chips (MPV?) : This is more specific then Search-based
compilation. Just to get an idea of what the chips are really
capable of.