Steps to Creating a Feature Matcher

• IMPORTANT: Get a new feature ID by the process discussed here.
• Create new files which over ride the functions in the abstract base classes. By doing this you automatically comply with the API. You should call your files something that resembles their function, e.g. HistogramRGB.c. Your classes should publicly inherit from FeatureVector and from ImageProcessor. To load and save your feature vector look at the way it's done in other modules, using the MemoryBlobWriter class.
• Update: You should integrate your code with the fvg program which allows you to generate and compare features from your code. The reasons for doing this are:
  1. It ensures your feature matcher and generator work correctly and allows you to test them.
  2. It ensures your feature matcher is Artiste API compliant.
  3. It keeps the fvg program up-to-date and makes it more powerful.

  Should this be happening concurrently you should integrate your feature matcher in isolation first, to fix any integration bugs. Once you are sure it's working the code maintainer will integrate it into the master copy of fvg.

  To integrate into fvg you need to follow these instructions. First of all, open convert.cpp in your copy of the fvg code directory. Insert into the code the relevant header file inclusion, and the relevant string compare function. This function converts the commandline into a FeatureVector and a ImageProcessor. For this code to work you will need to #define a string to compare against called NAME_xxx, where xxx is your feature vectors name - the one you used in the strcmp. This is defined in fvg.h in the include directory.

  That's it! Much easier than fvx was.

  The following is for archive only:

  You should integrate your code with the fvx program which allows you to generate and compare single feature vectors. (see below about concurrent development). You need to add new cases for your particular feature vector for the generate, compare and print attributes functionality. You should ensure that you update the help text located at the bottom of the code. Make sure you've instantiated a global (eugh!) instance of your object at the top of fvx.c. You will also have to update the Makefiles to include your function, by adding xxx.o to the OBJS and OBJSDIR variables near the top of the file, and adding its own compile command in the same manner as the previous features in the file.

• You should integrate your new feature generator and matcher into the Multiscale. program which allows you to compare many feature vector objects, for two reasons:
  1. It ensures that your feature matcher will work along with other feature matchers in the Artiste system, or a system which adheres to the Artiste API.
  2. Allows you to test your code against other feature matchers.
  3. It keeps the Multiscale program up-to-date, and makes it more powerful.
  Should this be happening concurrently with other developers you should integrate your feature matcher in isolation first, to fix any integration bugs. Once you are sure it's working the code maintainer (currently David Dupplaw) should be notified, who will gather the features which are being concurrently developed into one single instance of Multiscale.

  Update: The following has been updated!! Now you no longer need to implement a separate function for each of your modules. DoItRetrieval deals with the matching of all API compliant modules. After adding a new line into the list box in MultiScaleGui.c you should simply add a new call to this for the appropriate result of the GoButtonCallback. The makefile will link in the latest image processing library, and it should all just work. Multiscale still requires you to generate the feature vectors before hand, however, they are now in different directories. They should be within ReferenceImages and QueryImages and the lists are in ReferenceImages.db and QueryImages.db.

  The following is for archive only:

  Updating the Multiscale code is a simple, if somewhat ugly, task. In Multiscale.c, you need to copy a template of a function from, say, DoItHistogramRGB and change the instances of it's specific classes with your own. Yes, it's nasty, but updating it could take considerable time. You then need to add a procedure call to your new function, DoItXxx, in the GUI callback function, GoButtonCallback. Just copy a previous line and insert your function. Now you just need to add your feature vector to the pull down list box which is defined in MultiscaleGUI.c near the top of the code. Again, just copy a previous line, substituting your own text. Make sure you've instantiated a global (eugh!) instance of your object at the top of Multiscale.c.

  Remember: Mutliscale doesn't generate feature vectors. To run Mutliscale you need to first generate all the feature vectors with fvx into two directories (along with the original images): CCVedImages, and QueryImages, which is the default location for these. Then you need to generate the files CCVDatabase.db and QueryDatabase.db which contain the filenames of the original images (less the pathname), preceded by the number in the file. These are required for Multiscale to work.
  Much of this process could do with re-working. See "Stuff that'd be nice to do" for more details of things to do.

• Once your feature matcher has been verified to be working in a version of Multiscale, it can be passed to IT Innovations for integration into an Artiste release, if this is the final destination. This will be done by the code maintainer.