Image Morphing Project

1
Image Morphing Project

• Parallel Computing
• Mohammad S Rahim
• shahed_at_ece.gatech.edu
• March 18,2002.

2
Introduction

• Matt
• Paul
• They are 2 ECE graduate students I am image
  morphing.

3
Mesh Warping Steps

• Identify control points for each picture
• Place them in mesh matrices
• Iterate through each intermediary frame
• Find intermediary mesh for each frame
• Get the color mappings for this new mesh from
  each picture
• Do a weighted average of the 2 new pictures
  formed
• That is the new image for this intermediary frame

4
Mesh Warping Steps

• Identify control points for each picture
• Place them in mesh matrices
• Iterate through each intermediary frame
• Find intermediary mesh for each frame
• Get the color mappings for this new mesh from
  each picture
• Do a weighted average of the 2 new pictures
  formed
• That is the new image for this intermediary frame

5
Control Points

• Wrote the files into bitmaps
• Placed white dots where the control points should
  be
• Searched for differences from original picture
• Formed a matrix of the positions of the
  differences

6
Mesh Warping Steps

• Identify control points for each picture
• Place them in mesh matrices
• Iterate through each intermediary frame
• Find intermediary mesh for each frame
• Get the color mappings for this new mesh from
  each picture
• Do a weighted average of the 2 new pictures
  formed
• That is the new image for this intermediary frame

7
Mesh Matrices

• 9x9 matrix of the control points is created
• Lines between adjacent points in the matrix are
  shown here

8
Mesh Warping Steps

• Identify control points for each picture
• Place them in mesh matrices
• Iterate through each intermediary frame
• Find intermediary mesh for each frame
• Get the color mappings for this new mesh from
  each picture
• Do a weighted average of the 2 new pictures
  formed
• That is the new image for this intermediary frame

9
Intermediary Mesh Calculation
10
Mesh Warping Steps

• Identify control points for each picture
• Place them in mesh matrices
• Iterate through each intermediary frame
• Find intermediary mesh for each frame
• Get the color mappings for this new mesh from
  each picture
• Do a weighted average of the 2 new pictures
  formed
• That is the new image for this intermediary frame

11
Intermediary Mesh Calculation
12
Warp Individual Pictures

• The first picture shows the middle frame from
  Matt
• The second picture shows the middle frame from
  Paul

13
Middle Frame

• Here is the weighted average of the two images.

14
Compare
15
The Movie
16
Parallelizing

• Each frame can be computed on a different MPI
  machine like in a master-worker level system.
• But use OpenMP for the bulk of the computation
  within the frame.
• The main for loop in the problem is getting the
  color mappings for new mesh from each picture,
  (OpenMP should be used to handle that).
• The above loop is repeated at least 64 times,
  once for each mesh region. (It may be good idea
  to assign each region to a different processor)
  ltload balancing!?!gt

17
Serial Optimizations

• Matrixop.H
• 4x4, 4x2, 2x2 matrix multiplication
• 4x4 matrix inversion
• Of course addition and subtraction
• Computing regions in one loop run
• Very smart inpolygon function
• Not accessing disk in the calculation intensive
  process
• Pipe image saves to a separate BMP writer process
• 5 minutes (?) to 39.1136 sec for serial code
• Most of this is initialization time

18
OpenMP code

• Initialization is completely serial, but it has
  to be done only once
• Main portion of the code is the computation of
  the color maps region by region 64 regions
  (8x8)
• Many ways to parallelize this
• Parallelize the matrix calculations (not
  efficient enough)
• Courser grain, parallelize the regions
• Turns out it helps load balancing also
• Fairly linear speedup curve

19
(No Transcript)
20
SMP times

• Since initialization was not repeated
• Serial time 13.4s
• 8 proc time 2.59s
• So if we had 2.6 sec buffering time, we could
  display the morph real-time at a frame rate of
  1/2.6s 0.385fps!

21
Next Step

• Use distributed memory clusters JEDI.
• That potentially gives us 16x8 processors.
• MPI controls the application so it does not get
  in the way of others work.
• So dont get mad!
• Master- worker architecture.

22
(No Transcript)
23
Time

• Sysinfo gives load3. Number of busy
  msec/minute in the last 1 minute, 5 minutes, and
  15 minutes.
• It is easy to calculated load avg from there.
• I cut off work at 4.0 or higher.

24
Problems faced

• Mpicc does not accept mp argument.
• I could not find the omp libraries to link them.

25
My Solution

• So I compiled the MPI code and OpenMP code
  separately, and communicated between them using 2
  named pipes (FIFO).
• Communication cost of a named pipe is minimal.
• And since my OpenMP code was more or less
  untouched, my through-put is the same.

26
Result

• (0.385 (lt16)) lt 6.15 fps.
• We can actually see something without falling
  asleep.

27
I love seeing this video!