You need to create a global representation of your local features so that you can feed your data to SVMs. One of the most popular approaches for this task is bag-of-words(features). vlfeat has an excellent demo/example for this. You can check this code from vlfeat website.
For your particular case, you need arrange your training/testing data in Caltech-101 like data directories:
- Letter 1
- Image 1
- Image 2
- Image 3
- Image 4
- ...
- Letter 2
- ...
- Letter 3
- ...
Then you need to adjust following configuration settings for your case:
conf.numTrain = 15 ;
conf.numTest = 15 ;
conf.numClasses = 102 ;
This demo uses SIFT as local features, but you can change it to HOG afterwards.