Parallel Analysis of Fischlar Digital Video

1
Parallel Analysis of Fischlar Digital Video
IAHPC Meeting, November 29th 2002
Karl Podesta karl.podesta_at_computing.dcu.ie Scho
ol of Computing
2
(No Transcript)
3
Methods for Detecting Shot Changes
- Colour Histograms Compare the Colour
Histograms of successive frames Register a cut
on big changes
- Edge Detection Greyscale, SOBEL filter,
Dilate Compare and cut
- MPEG Macroblock Done on the MPEG file
(compressed)
4
The Motivation for Parallel
There are LOTS of Frames! . 4000 frames for a
2-min video
For each frame () Extract a JPEG, () Get the
Colour Histogram, () Compare with previous
histogram register cut
5
The Initial Solution
Simple Data Decomposition (MPEG)
Split the MPEG video across the number of compute
nodes Let nodes churn away on the data Border
Communicate histograms for comparison
6
Issues for the Implementation

• MPEG Byte Splitting is problematic
• Loss of Shot Accuracy
• GOP, a unit between I-frames
• Split by GOP!

• Managing Large or Streaming Video
• Use an Optimal Chunk or Window to step through
  data
• About 250k on a basic Beowulf

7
A performance improvement
Sun Enterprise 4500 4GB RAM 4 x Ultrasparc
II (OLD PLATFORM) VS Linux Beowulf 20 nodes x
i386 P166Mhz (NEW PLATFORM)
Good speedup, approx 5 times faster 2 minute
piece of video - 2m52s on Sun, 35s on Beowulf
8
Salvaging time and solutions
Extra Room for additional methods, accurate
detection for thinking about more capability
(Object Identification, Homer Detector)
Not to mention better Parallel Implementation Gen
etic Algorithms, Hidden Markov Models, Spatial
Decomposition - all suitable for MPEG
9
Enabling Computing

• The Fischlar System
• Shot Scene Detection a tiny part
• Parallel Video Analysis Small but useful
• Can HPC help everyone?