# Last edited on 2003-03-03 10:01:16 by stolfi

DATASET DESCRIPTION

  Fragment preparation
  
    This directory contains the input data for the test "glazed-1".
    
    The test objects were 9 glazed wall tiles, each 19.5 cm square.
    The ceramic was fine grained, chocolate brown, with a smooth
    fracture. The glaze was opaque, light turquoise, a fraction of mm
    thick. The scanned (enameled) surface had a fine bumpy texture ???
    
    All tiles but one were shattered into small fragments (max 5--6 cm
    diameter) by being placed on a newspaper-covered concrete surface,
    glazed face down, and being repeatedly hit with a steel bar. The
    last tile was broken into 6 large pieces (0000 through 0005) by
    stepping onto it while one edge was resting on the steel bar.
    
    The fragments were scanned on an IBM ??? flatbed scanner, driven
    by the Adobe EasyPhoto software that comes with the product. The scanning
    resolution was set to 600 dpi (which actually turns out to be 593
    dpi, or 0.0428 mm/pixel). We were unable to prevent that crappy
    brain-damaged software from using JPEG format at unknown quality
    setting, and performing automatic brightness and constrast
    normalization --- with widely different settings for each scan.
    
    The JPEG image for fragment batch "b" was accidentally lost and
    had to be reconstructed from a normalized ".pgm" version.

    Some fragment images (notably 0007, 0013, 0018, 0305, 0307, 0308,
    0314), were damaged because of uneven illumination by the scanner,
    which foiled the simple threshhold-based segmentation program.
    Fragment 0400, a thin sliver, was scanned on its side; fragment
    0037 was partly clipped during scanning. Also, original fragment
    edges were badly reproduced because the original tiles had a 2mm
    wide chamfer all around.
    
    On some fragments, the numeric label ended up touching the 
    fragment outline, which at first caused problems for the
    outline extraction software. The offending batches were retouched
    by hand to correct the problem.

FILES

    For each batch X in {"a" .. "j"}, the directory "data/batches/X"
    contains 
    
      image-raw-orig.jpg     scanner output
      image-raw-edited.png   ditto, retouched by hand 

>>>STOPPED HERE - The following files are obsolete and must be recreated.

      sample-raw.pgm         OBSOLETE - RECREATE

      image-norm.pgm.gz      OBSOLETE - RECREATE
      sample-raw.hist        OBSOLETE - RECREATE
      set-gray-levels.csh    OBSOLETE - RECREATE
      
    The directory "data/fragments" contains one subdirectory 
    for each fragment, whose name "NNNN" is the fragment number.
    For each fragment, the file "image.pgm.gz" contains the 
    isolated fragment, against a black background. 
    The extracted outline, sampled at irregular intervals, is "r000.flc".
    
    The directory "data/multiscale" also contains one sub-directory
    "NNNN" for each fragment. The file "NNNN/f000.flc" contains the
    outline of fragment NNNN, resampled with uniform step
    "f000.lambda"/4. The file "f000.lbl" contains the corresponding
    label values (length along the curve). Then, for each non-zero
    BBB, the file "fBBB.flc" is the output of PZFilter for scale BBB,
    namely for lambda = "fBBB.lambda". The files "fBBB.fla" and
    "fBBB.flv" are the estimated velocity and acceleration at each
    sample. "fBBB.fcv" is the estimated curvature at each sample, and
    "fBBB.lbl" are the corresponding label values (which can be
    matched to the values in "f000.lbl").
    
    The directory "data/nonfractal" contains data on "straight" parts
    of the outlines, which are unsuitable for multiscale matching.
    
    
    
      
Rerun: 
  normalization
  splitting
  renaming
  boundary extraction
  printing
  filtering
  curvature computation
  curvature histogram
  straight segments
  true candidates
  information contents
  ...