Abstract
Artificial Intelligence (AI)-based emotion recognition using various kinds of data has attracted vast attention in recent years. Impressive results have been achieved, but only recently the influence of the training data with its potential biases and variations in annotation quality are discussed. Still, the majority of the research literature focuses on improving machine learning techniques and model performance using single data sets. Literature on the impact of training data remains scarce. Therefore, in this paper we investigate the influence of the training data on the accuracy of recognizing emotional states in facial expressions by a comparative evaluation, using multiple established facial image databases. Results reveal inconsistencies in the data annotations as well as ambiguities in the emotional states expressed. Thus, they allow to critically discuss data quality of the training data, contributing to a more in-depth understanding of previous emotion recognition approaches, and improving the design of more transparent AI solutions.
Users
Please
log in to take part in the discussion (add own reviews or comments).