For bounding boxes that are false positives, it would be nice if the letter classifier could tell us if it doesn't think there's a letter there. Or, we just keep it the current way where if the distribution of likelihoods for each letter is uniform then we assume it's less likely to be a target.
For bounding boxes that are false positives, it would be nice if the letter classifier could tell us if it doesn't think there's a letter there. Or, we just keep it the current way where if the distribution of likelihoods for each letter is uniform then we assume it's less likely to be a target.