Paper accepted at ACII 2025. A collaborative work of Seeing Machines and University of Canberra for improved valence estimation. Congratulations to all the co-authors!

Paper titled “Diversity benefits: Integrating Datasets for Improved Valence Estimation” accepted at ACII 2025.

Abstract:

Robust and generalisable emotion inference (EI) is critical due to its wide applicability from mental health assessment to driver monitoring. Affective training data quality impacts EI performance, and building robust EI models involves multiple challenges. Emotions are subjective, context dependent, and influenced by cultural and demographic variations. Emotional data annotation is both tedious and onerous as humans are better at rating in relative than absolute terms. Given the (1) scarcity of large-scale affective datasets and (2) non-standardised affective labelling procedures across datasets, learning from heterogeneous data can enable robust EI. We discuss the impact of training EI models upon integrating data (a) captured in-the-wild vis-`a-vis in controlled conditions, and (b) depicting acted versus naturalistic emotions. Experiments with the AffectNet, EMMA and RECOLA datasets reveal that careful addition of cross-dataset samples can improve valence estimation CCC and PCC by more than 11%.