Emotion recognition is shaping the future of mental health, behavior understanding, and HCI. While speech and facial expressions have dominated research, physiological signals (EDA, PPG) give a more authentic window into our emotions — harder to mask and increasingly accessible through wearables. What’s missing? A unified benchmark. Current datasets are too small and fragmented, holding back progress toward truly generalizable models.
Our work addresses this gap by introducing FEEL, the first unified benchmarking framework for physiological signal–based emotion recognition.
Key contributions:
1) A harmonization strategy to standardize heterogeneous datasets.
2) A novel fine-tuning method for contrastive language-signal pretraining (CLSP).
3) A comprehensive cross-dataset analysis, showing how gender, age, environment, device type, and labeling method affect generalization.
4) Benchmarking 19 publicly available datasets, covering a wide range of settings, devices, and labeling schemes.
By moving beyond isolated dataset studies, FEEL provides a systematic foundation for evaluating and developing robust, transferable models in physiological emotion recognition.
Link