Analysis of the Calibration of Handwriting Text Recognition Models

Current neural approaches for Handwritten Text Recognition (HTR) have proven to be successful in many settings, but their performance can be unpredictable when facing new data. In this context, it is essential to correctly estimate an approximate error of the target predictions. To achieve this, the model must be well calibrated, meaning that the confidence values are sufficiently representative of the expected accuracy. Calibration is a crucial aspect in practical applications of HTR, but the topic remains overly underexplored. In this paper, we fill this gap by studying calibration in state-of-the-art HTR models, along with specific techniques for this purpose. We perform thorough experiments on several datasets, both in a classic setting and in cross-collection scenarios. Our results report interesting conclusions about the calibration of HTR, highlighting their strengths, weaknesses, and the extent to which the considered strategies improve results.