Affective computing has experienced substantial advancements in recognizing emotions through image and facial expression analysis. However, the incorporation of physiological data remains constrained. Emotion recognition with physiological data shows promising results in controlled experiments but lacks generalization to real-world settings. To address this, we present G-REx, a dataset for real-world affective computing. We collected physiological data (photoplethysmography and electrodermal activity) using a wrist-worn device during long-duration movie sessions. Emotion annotations were retrospectively performed on segments with elevated physiological responses. The dataset includes over 31 movie sessions, totaling 380 h+ of data from 190+ subjects. The data were collected in a group setting, which can give further context to emotion recognition systems. Our setup aims to be easily replicable in any real-life scenario, facilitating the collection of large datasets for novel affective computing systems.

Emotion recognition systems are typically trained to classify a given psychophysiological state into emotion categories. Current platforms for emotion ground-truth collection show limitations for real-world scenarios of long-duration content (e.g. >10 minutes), namely: 1) Real-time annotation tools are distracting and become exhausting; 2) Perform retrospective annotation of the whole content in bulk (providing highly coarse annotations); or 3) Are used by external experts (depending on the number of annotators and their subjective experience). We explore a novel approach, the EmotiphAI Annotator, that allows undisturbed content visualisation and simplifies the annotation process by using segmentation algorithms that select brief clips for emotional annotation retrospectively. We compare three methods for content segmentation based on physiological data (Electrodermal Activity (EDA), emotion-based), scene (time-based), and random (control) selection. The EmotiphAI Annotator attained a B+ System Usability Scale score and low-average mental workload as per the NASA Task Load Index (40%). The reliability of the self-report was analysed by the inter-rater agreement (STD < 0.75), coherence across time segmentation methods (STD < 0.17), comparison against the state-of-the-art ground truth (STD < 0.7), and correlation to EDA (>0.3 to 0.8), where the EDA-based method obtained the overall best performance.