Quantitative assessment of performance is not generally yielded by the authors of the
methods; when it is given, this is on a reduced set of documents. As for all segmentation
methods, ground truth data are harder to obtain than for classification methods. For instance
the ground truth for the real baseline may be hard to assess. Text line segmentation is often a
step in the recognition algorithmand the segmentation task is not evaluated in isolation. To
date, no general study has been carried out to compare the different methods. Text line
representations differ and methods are generally tuned to a class of documents.