Open mz2 opened 10 years ago
I've built a training and test datasets for the Greek alphabet. It looks like 64 per symbol is way above what's needed for decent classification performance at least in this alphabet (5% overall error rate with 220 examples in total of the full alphabet).
Actually, 5% error rate was a bit optimistic, that was an out of bag error estimate from the classifier. With an independent test suite with 2 x of each character the achieved precision is 78%.
Build sets of training examples for proofreading markers using a mouse / Magic Trackpad and symbol sizes of the size applicable to the task (e.g. proofreading symbols are quite small, LaTeX symbols you can draw a bit larger, basic geometric shapes can vary in size and skewedness).
n for each symbol type should be ~64 per symbol for each person drawing the symbol.