Positive and negative samples for haartraining

edit

Don't start with glasses. They are transparent and thus difficult. Start with something simple. Mugs would be better. Or some fruits like apples or bananas. Take more than 17 images, say 50-100, of your objects, each slightly different, each on different background, from different angle, preferably use many bananas or apples. Next take a number of images containing possible backgrounds for your objects WITHOUT these objects. For example 10 images of your kitchen. During training specify much more negatives than you really have images - the training function will randomly cut out these negatives from your images - so your negatives must be relatively big, don't cut them. From 10 high resolution images of your kitchen, the training function will easily generate thousands of negative samples. Experiment a bit. After some training you should be able to get a classifier able to recognize apples or bananas in your kitchen. However, it will not be robust in all possible environments. For good performance in almost any conditions you would need hundreds if not thousands of positive samples and thousands or tens of thousands of negative ones.