----------------------------------------------------------------------------
README file for Temporal texture directory

  This directory contains image sequences of temporal textures,
such as wavy water and smoke.

These sequences were used in the two publications:

Temporal Texture Modeling 
     Martin Szummer and Rosalind W. Picard 
     IEEE International Conference on Image Processing (ICIP 1996).
     Lausanne, Switzerland. 
     ftp://whitechapel.media.mit.edu/pub/tech-reports/TR-381.ps.Z

Temporal Texture Modeling 
     Martin Szummer 
     M.Eng. Thesis, MIT, Sept 1995.
     Also appeared as MIT Media Lab Perceptual Computing TR #346 
     ftp://whitechapel.media.mit.edu/pub/tech-reports/TR-346.ps.Z

Also see http://www.media.mit.edu/~szummer/
         http://www-white.media.mit.edu/~szummer/icip-96/
----------------------------------------------------------------------------
Contents:
 Most of the sequence names are self-explanatory.
 Some sequences occur several times; some sequences may be close-ups 
 of particular parts of the image.
 The sequences have resolutions of about 170x115 and 120 frames,
   but some are different.
 Most sequences are grayscale only, except for fire.rgb.

 raw/      - directory contains original raw texture files
    boil*  - contains boiling water
    escalator - moving stairs
    fire      
    fire.rgb  - color fire
    flags     - flags flapping in the wind
    fountain
    laundry   - landromat spinning clothes
    plastic   - plastic sheet waving in the wind
    river-far - wavy water from far
    river     - wavy water
    shower    - shower droplets on glass
    smoke     - smoke (or steam?) rising from chimney
    steam     - steam from manhole
    stripes   - American flag waving in the wind
    toilet    - spiralling water
    trees     - trees waving in the wind
 
 synthesis/   - directory contains synthesized temporal textures
                NOTE: most of these textures have been synthesized
                based on unsharp masked versions of the images
                (ie subtract off a frame-by-frame median filtered version of 
                 the sequence).  Therefore the intensity distribution
                differs from the raw files.

    The synthesized files all used the autoregressive model.
    The neighborhoods used were:
      *.C4      - causal sphere of radius 4  (128 parameters)
      *.R10     - causal cube with side 10   (1270 parameters)

How the sequences were acquired:

  The data consists of commonly occurring temporal textures The
  particular sequences were selected to get interesting, every-day
  indoor and outdoor textures with different motions.  With a few
  exceptions, the chosen textures appear to be wide sense stationary and
  amenable to statistical texture techniques.  The data was acquired by
  me to avoid copyright issues.

  The image sequences of temporal textures were recorded using a Hi-8
  video camera.  A tripod was used to avoid camera motion.  The
  sequences were digitized at a resolution of 720x487, Y-B-R (8-4-4 bit)
  color and 30 frames per second onto D1 digital format.  Next, the
  sequences were cropped to only contain temporal textures, then
  deinterlaced, low-pass filtered and subsampled.  Only the luminance
  information was used.  The final resolution in all cases was about
  170x115 and the length was 120 frames (corresponding to 4 seconds).
  Thus, each sequence contains about 2 million data points.

----------------------------------------------------------------------------
INSTRUCTIONS for downloading the sequences:

 1) ftp the data files in binary mode.
       . please download the files at night to minimize load on the network
       . if the ftp connection is unreliable, use an ftp client that
         can recover from lost connections.  (ie. ncftp or libftp)
 2) The files with .tgz extension are tarred and gzipped (GNU zip),
     (on an HP 700 series workstation).
    On most UNIX systems, the following command should extract one file
       gunzip -c filename.tgz | tar xf -
 3) Each extracted sequence now resides in a separate directory.
    The directory contains two files:
         descriptor - stores sequence dimensions and history 
         data       - raw binary data, without any headers.
                      not compressed in order to avoid artifacts

 Structure of descriptor file:
  _data_sets   - number of individual frames
  _channels    - number of color channels
  _dimensions  - two numbers, specifying height first, then width
  _data_type   - unsigned_1 corresponds to "unsigned char" in C
 
 Structure of data file:
  Frames are written one by one.
  Color channels (with _channels equal to 3) are written in
   non-interleaved order, ie. first all R pixels for a frame, then all G
   and B.

 To view the sequences, you either need to have a program that views
 raw sequences directly, or convert the sequences to some other format.
 I used the Utah Raster Toolkit's program rawtorle and then getx11.
 The toolkit is at:  http://www.cs.utah.edu/gdc/projects/urt/
 (if this URL is broken, use a search engine and look for "Utah Raster Toolkit").

----------------------------------------------------------------------------
 Permission: 
    You are free to use these sequences for research, and include them in
your publications as long as you acknowledge the original source.
I would also be very interested in learning about your research!

- Martin Szummer  12/17/96
  MIT Media Lab
  szummer.NOSPAM@media.mit.edu     (remove NOSPAM suffix before sending)