Adds two more examples of croped word recognition, and show how to set the beam size parameter for small lexicon.

OCRBeamSearchDecoder code refactor: fixes most of the TODOs comments, cleans-up code, and fix a bug in score_segmentation(...) member function.

Add benchmark code for the Chars74k dataset. Using the CNN character classifier reaches 75% and 84% accuracy for case-sensitive and case-insensitive recognition respectively.

Wrap the ROISelector class

Add benchmark for ICDAR2015 dataset using OCRTesseract and ERFilter classes. Gives word spotting f-score 0.642082 with strongly contextualized lexicon (100 words).

Report the total f1 score (0.37) and not the mean-f1. This is in accordance with the standard SVT evaluation protocol and allows for comparison with other published results.