# Last edited on 2011-04-30 04:32:08 by stolfi

UNPACKING

  Used two commands:
  
    ffmpeg -i video.mpg frames-a/%08d.jpg
    
    ffmpeg -i video.mpg -vcodec png frames-b/%08d.png
    
  Obtained 196 frames. Deleted frames 00000001 through 00000037
  and 00000154 through 00000196 (camera out of water).
  
COMPARING

  Comparing the two results:
  
    convert frames-a/00000130.jpg .a.ppm
    convert frames-b/00000130.png .b.ppm
    pnmxarith -subtract -scale 10 -offset 0.5 .a.ppm .b.ppm > .diff.ppm
    display .diff.ppm

  There is a difference, it seems. Trying to compare visually:
    
    display -filter box -resize '200%' .a.ppm .b.ppm &
    
  The frames-b output looks better; the frames-a has visible 8x8 JPEG
  blockiness. Whether it is more accurate, or just smoothed out, I do 
  not know.  Anyway:
  
    rm -rf frames-a
    
COMPOSITING

  Trying to align successive frames by hand. Let's pick some frames 
  from the middle of the movie 
  
    mkdir frames-p
    frames=( 0000009{0,1,2,3,4,5,6,7,8,9} )

    for f in ${frames[@]} ; do 
      bfile="frames-b/${f}.png"
      pfile="frames-p/${f}.ppm"
      convert ${bfile} ${pfile}
    done
    
    ( cd frames-p && display -title '%f' ${frames[@]/%/.ppm} ) &
    
  Created for each frame ${f} a file frames-p/${f}.pts
  containing the pixel coords of 4 referece points.
    
  Pick a reference frame ${ref} and apply a projective map
  correction to all frames to match is reference points:
  
    ref=00000095
    rpts="frames-p/${ref}.pts"
    for f in ${frames[@]} ; do 
      pfile="frames-p/${f}.ppm"
      ppts="frames-p/${f}.pts"
      ofile="frames-p/${f}-a.ppm"
      pnmprojmap \
          -points `cat ${ppts}` `cat ${rpts}` \
          -interpolate 0 \
          -oSize 900 600 \
          -oOrg 90 60 \
          -maxval 65535 \
          ${pfile} \
        > ${ofile}
    done
      
    ( cd frames-p && display -title '%f' ${frames[@]/%/-a.ppm} ) &
    
  The alignments were not so good, because (1) the ref points were not
  very accurate, (2) the second reference point may not be coplanar
  with the other three (which lie at the top of the fuel assemblies),
  and, mainly, (3) there is much image deformation by thermal
  gradients in the water.
  
  I must write a multi-image alignment program that combines a
  projective map with a low-frequency harmonic series distortion.

  Let's try mixing the corrected images anyway:
  
    pnmxarith -mix 0.50000 0.50000 frames-p/00000090-a.ppm frames-p/00000091-a.ppm > .a01.ppm
    pnmxarith -mix 0.50000 0.50000 frames-p/00000092-a.ppm frames-p/00000093-a.ppm > .a23.ppm
    pnmxarith -mix 0.50000 0.50000 frames-p/00000095-a.ppm frames-p/00000096-a.ppm > .a56.ppm
    pnmxarith -mix 0.50000 0.50000 frames-p/00000097-a.ppm frames-p/00000098-a.ppm > .a78.ppm
   
    pnmxarith -mix 0.66667 0.33333 .a23.ppm frames-p/00000094-a.ppm > .a234.ppm
    pnmxarith -mix 0.66667 0.33333 .a78.ppm frames-p/00000099-a.ppm > .a789.ppm
    
    pnmxarith -mix 0.40000 0.60000 .a01.ppm .a234.ppm > .a01234.ppm
    pnmxarith -mix 0.50000 0.50000 .a56.ppm .a789.ppm > .a56789.ppm

    pnmxarith -mix 0.50000 0.50000 .a01234.ppm .a56789.ppm > .atot.ppm
    
    convert .atot.ppm .atot.png
    
    gimp .atot.png &
    
  The result is not bad! Adjusted channel curves with gimp, cropped and saved as 
  "frames-b/avg-090-099-c.png"