Citation

Hew, Zhong Ken and Wong, Lai Kuan and Chan, Chee Seng (2026) Beyond illusions of competence: Revisiting zero-shot learning in emotion recognition with FEAr dataset. Signal Processing: Image Communication, 148. p. 117658. ISSN 09235965

Text Beyond illusions of competence_ Revisiting zero-shot learning in emotion recognition with FEAr dataset.pdf - Published Version Restricted to Repository staff only Download (4MB)

Abstract

This paper investigates the efficacy of Zero-Shot Learning (ZSL) in the domain of emotion recognition, uncovering three significant shortcomings: (i) deceptive performance, (ii) a lack of standardized metrics and (iii) a deficiency in robust emotion semantic attributes. Our analysis shows that current ZSL methods suffer from misleading performance evaluations due to inconsistent dataset splits, which inflate results and obscure the true capabilities of ZSL models in emotion recognition. Moreover, the absence of standardized evaluation metrics exacerbates this issue, making it challenging to compare and validate the effectiveness of different ZSL models accurately. Moreover, the lack of robust semantic attributes for emotions hampers ZSL performance in emotion recognition, making it difficult to align these attributes with visual features. To address these challenges, we propose a new framework, namely Fair Evaluation for Affective Recognition (FEAr) alongside the FEAr-accuracy Harmonization (FaH) evaluation metric as innovative solutions to these problems. The FEAr framework establishes standardized, valence-balanced split configurations for benchmark emotion datasets, enabling consistent and reproducible evaluation in zero-shot learning. Meanwhile, the FaH metric offers a robust framework for evaluating ZSL models, enabling reliable cross-study comparisons. Furthermore, the CLIPAffective Attributes Generator (CLIP-AAG) is proposed to generate emotional semantic attributes that are more effectively aligned with the corresponding emotional visual information. This study aims to foster a paradigm shift towards transparent, comparable, and replicable research practices in the field of emotion recognition. The dataset and code associated with this work are available at https://github.com/HewZK2000/CLIP-AAG.git

Item Type:	Article
Uncontrolled Keywords:	Zero-shot learning, Image emotion recognition
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75.5-76.95 Electronic computers. Computer science
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	03 Sep 2026 02:13
Last Modified:	03 Sep 2026 02:13
URII:	http://shdl.mmu.edu.my/id/eprint/16576

Downloads

Edit (login required)