MAID: Multimodal affective imagery dataset for detecting and predicting emotions across tasks

Recent years have shown rapid advancement in neurophysiological emotion detection. However, the nature of such data makes it unclear whether reports of very high accuracy are due to the genuine ability to detect affective states, or due to detecting more superficial differences between stimuli. This may explain why EEG-based emotion detection accuracy drastically declines when stimulus features change between contexts or tasks. We argue the field is in need of a dataset that keeps perceptual differences constant while testing the detection of emotion via multiple sensor modalities to predict transfer across tasks. We present an open multimodal, affective imagery dataset (N=44) with measures of EEG, facial EMG, EDA, ECG, accelerometer data and EOG. In task 1, participants imagined having low or high arousal and feeling negative or positive, while in task 2, they imagined being in situations that would induce similar emotions. Benchmark tests \redwas{showed both within-task emotion detection accuracy and cross-task prediction accuracy. Indeed, only facial EMG, and therefore emotional expression, showed clear transferability, which has important theoretical and practical repercussions. To verify whether the problem is due to processing steps or choice of algorithm, we openly release our multimodal dataset as a challenge for the affective computing community to detect emotions without stimulus-related activity. Ultimately, we argue, the neurophysiological computing target is not about detecting if a picture looks emotional, but whether a person is, even if there is no picture at all.