Assignment 2: word zone hypotheses

The page may be slightly changed in the future. Be sure to refresh (F5) right before you start.

Goal

The goal of this assignment is to make a generator for word zone hypotheses.

A word zone hypothesis is a set of coordinates defining a region (zone) that possibly contains a word. This can be the first step of a word-based handwriting recognizer. To save you some hassle, the line zones are provided.

Assignment

Make a word zone hypothesis generator. Given an image and a file describing the line zones, output a new file describing an abundance of word zones.

The command to call the program should be like this:
python zonehypo.py input.ppm input.words output.words
where:

• input.ppm is a complete page, such as NL_HaNa_H2_7823_xxxx.ppm;
• input.words contains the line zones (but no word zones or no usable word zones);
• output.words will be written and contain word zones without transcription.

The input and output files with extension .words have a format as defined here: .words file description. Reading and writing such files is made easy by the provided wordio.py.

Make sure that the C++ part of your code compiles with the command 'make'. This is required.

Submit your program using Nestor's Digital Dropbox.

Hints

• Copy the quickstart files:
  cp -r /home/student/hwr/assignment2 ~/hwr/
cp -r /home/student/hwr/wordsfile-editor ~/hwr/
  Note: the files are described in the Appendix below.
• Have a look at the two train sets:
  • /home/student/hwr/lines-train
  • /home/student/hwr/words-train
  These directories contain XML files with file extension .words. The first one is similar to the secret dataset that will be used for testing; it has no <Word> tags. The second version includes the correct word zones, including correct annotation. (These were manually created, students of 2007 and 2008 worked hard for it; you are spared this year.)
• The program wordsfile-editor/wordsfile-editor.py can be used to inspect and/or modify .words files:
  python wordsfile-editor.py imagefile wordsfile.words
  Try it on a file from lines-train and a file from words-train. If you make a major improvement to a .words file, please send it to me because it might be useful.
  Note: by default, the .words files are write-protected. This can be changed by using chmod u+w *.words
• Try the word zone generator example:
  python generate_wordzones.py input.ppm input.words output.words
  It reads an input image and input .words file with line zones, and writes an output .words file with an abundance of random word zones.
• Try the connected-components example:
  First use an image manipulation program (such as Gimp) to binarize an image (make it black-and white by thresholding) and save it as raw .pgm. Then use:
  python example-cocos.py input.pgm output.pgm
  It will crop a black connected component and show some information.
• You can use the program netpbm to shear an image, if you like.

Grading

• The generated zones will be matched with "true" (manually labeled) word zones. At least 60% of the true zones must be found (the coordinates are checked; a small error is tolerated).
• The less zone hypotheses are generated, the higher you score.

Appendix: Description of files in assignment2/