Next Generation Typeface Representations: Revisiting Parametric Fonts

Abstract

A number of methods for evaluating table structure recognition systems have been proposed in the literature, which have been used successfully for automatic and manual optimization of their respective algorithms. Unfortunately, the lack of standard, ground-truthed datasets coupled with the ambiguous nature of how humans interpret tabular data has made it difficult to compare the obtained results between different systems developed by different research groups.

With reference to these approaches, we describe our experiences in comparing our algorithm for table detection and structure recognition to another recently published system using a freely available dataset of 75 PDF documents. Based on examples from this dataset, we define several classes of errors and propose how they can be treated consistently to eliminate ambiguities and ensure the repeatability of the results and their comparability between different systems from different research groups.