Fine-grained emotion recognition can model the temporal dynamics of emotions, which is more precise than predicting one emotion retrospectively for an activity (e.g., video clip watching). Previous works require large amounts of continuously annotated data to train an accurate recognition model, however experiments to collect such large amounts of continuously annotated physiological signals are costly and time-consuming. To overcome this challenge, we propose an Emotion recognition algorithm based on Deep Siamese Networks (EmoDSN) which can rapidly converge on a small amount of training data, typically less than 10 samples per class (i.e., <10 shot). EmoDSN recognizes fine-grained valence and arousal (V-A) labels by maximizing the distance metric between signal segments with different V-A labels. We tested EmoDSN on three different datasets collected in three different environments: desktop, mobile and HMD-based virtual reality, respectively. The results from our experiments show that EmoDSN achieves promising results for both one-dimension binary (high/low V-A, 1D-2 C) and two-dimensional 5-class (four quadrants of V- A space + neutral, 2D-5 C) classification. We get an averaged accuracy of 76.04, 76.62 and 57.62% for 1D-2 C valence, 1D-2 C arousal, and 2D-5 C, respectively, by using only 5 shots of training data. Our experiments show that EmoDSN can achieve better results if we select training samples from the changing points of emotion or the ending moments of video watching.

While affective non-verbal communication between pedestrians and drivers has been shown to improve on-road safety and driving experiences, it remains a challenge to design driver assistance systems that can automatically capture these affective cues. In this early work, we identify users' emotional self-report responses towards commonly occurring pedestrian actions while crossing a road. We conducted a crowd-sourced web-based survey (N=91), where respondents with prior driving experience viewed videos of 25 pedestrian interaction scenarios selected from the JAAD (Joint Attention for Autonomous Driving) dataset, and thereafter provided valence and arousal self-reports. We found participants' emotion self-reports (especially valence) are strongly influenced by actions including hand waving, nodding, impolite hand gestures, and inattentive pedestrian(s) crossing while engaged with a phone. Our findings provide a first step towards designing in-vehicle empathic interfaces that can assist in driver emotion regulation during on-road interactions, where the identified pedestrian actions serve as future driver emotion induction stimuli.

Voice User Interfaces (VUIs) such as Alexa and Google Home that use human-like design cues are an increasingly popular means for accessing news. Self-disclosure in particular may be used to build relationships of trust with users who may reveal intimate details about themselves. This information can be (mis)used by algorithms to tailor and deliver partisan, critical news at the cost of journalistic ethics. In this position paper, we argue that self-disclosing VUIs may be beneficial to build trust with users and deliver news. We explain how a self-disclosing relationship may not only be a step towards human-like voice assistants but also aid in acceptance of, and exposure to different news events. We caution against obvious pitfalls from both a system and journalistic perspective and provide measures to address such concerns.

Automatically inferring drivers' emotions during driver-pedestrian interactions to improve road safety remains a challenge for designing in-vehicle, empathic interfaces. To that end, we carried out a lab-based study using a combination of camera and physiological sensors. We collected participants' (N=21) real-time, affective (emotion self-reports, heart rate, pupil diameter, skin conductance, and facial temperatures) responses towards non-verbal, pedestrian crossing videos from the Joint Attention for Autonomous Driving (JAAD) dataset. Our findings reveal that positive, non-verbal, pedestrian crossing actions in the videos elicit higher valence ratings from participants, while non-positive actions elicit higher arousal. Different pedestrian crossing actions in the videos also have a significant influence on participants' physiological signals (heart rate, pupil diameter, skin conductance) and facial temperatures. Our findings provide a first step toward enabling in-car empathic interfaces that draw on behavioural and physiological sensing to in situ infer driver emotions during non-verbal pedestrian interactions.

Social VR enables people to interact over distance with others in real-time. It allows remote people, typically represented as avatars, to communicate and perform activities together in a shared virtual environment, extending the capabilities of traditional social platforms like Facebook and Netflix. This paper explores the benefits and drawbacks provided by a lightweight and low-cost Social VR platform (SocialVR), in which users are captured by several cameras and reconstructed in real-time. In particular, the paper contributes with (1) the design and evaluation of an experimental protocol for Social VR experiences; (2) the report of a production workflow for this new type of media experiences; and (3) the results of experiments with both end-users (N = 15 pairs) and professionals (N = 22 companies) to evaluate the potential of the SocialVR platform. Results from the questionnaires and semi-structured interviews show that end-users rated positively towards the experiences provided by the SocialVR platform, which enabled them to sense emotions and communicate effortlessly. End-users perceived the photo-realistic experience of SocialVR similar to face-to-face scenarios and appreciated this new creative medium. From a commercial perspective, professionals confirmed the potential of this communication medium and encourage further research for the adoption of the platform in the commercial landscape.

Watching 360 videos using Virtual Reality (VR) head-mounted displays (HMDs) provides interactive and immersive experiences, where videos can evoke different emotions. Existing emotion self-report techniques within VR however are either retrospective or interrupt the immersive experience. To address this, we introduce the Continuous Physiological and Behavioral Emotion Annotation Dataset for 360 Videos (CEAP-360VR). We conducted a controlled study (N=32) where participants used a Vive Pro Eye HMD to watch eight validated affective 360 video clips, and annotated their valence and arousal (V-A) continuously. We collected (a) behavioral (head and eye movements; pupillometry) signals (b) physiological (heart rate, skin temperature, electrodermal activity) responses (c) momentary emotion self-reports (d) within-VR discrete emotion ratings (e) motion sickness, presence, and workload. We show the consistency of continuous annotation trajectories and verify their mean V-A annotations. We find high consistency between viewed 360 video regions across subjects, with higher consistency for eye than head movements. We furthermore run baseline classification experiments, where Random Forest classifiers with 2s segments show good accuracies for subject-independent models: 66.80% (V) and 64.26% (A) for binary classification; 49.92% (V) and 52.20% (A) for 3-class classification. Our open dataset allows further experiments with continuous emotion self-reports collected in 360 VR environments, which can enable automatic assessment of immersive Quality of Experience (QoE) andmomentary affective states.

Recognizing user emotions while they watch short-form videos anytime and anywhere is essential for facilitating video content customization and personalization. However, most works either classify a single emotion per video stimuli, or are restricted to static, desktop environments. To address this, we propose a correlation-based emotion recognition algorithm (CorrNet) to recognize the valence and arousal (V-A) of each instance (fine-grained segment of signals) using only wearable, physiological signals (e.g., electrodermal activity, heart rate). CorrNet takes advantage of features both inside each instance (intra-modality features) and between different instances for the same video stimuli (correlation-based features). We first test our approach on an indoor-desktop affect dataset (CASE), and thereafter on an outdoor-mobile affect dataset (MERCA) which we collected using a smart wristband and wearable eyetracker. Results show that for subject-independent binary classification (high-low), CorrNet yields promising recognition accuracies: 76.37% and 74.03% for V-A on CASE, and 70.29% and 68.15% for V-A on MERCA. Our findings show: (1) instance segment lengths between 1–4 s result in highest recognition accuracies (2) accuracies between laboratory-grade and wearable sensors are comparable, even under low sampling rates (≤64 Hz) (3) large amounts of neu-tral V-A labels, an artifact of continuous affect annotation, result in varied recognition performance.

We demonstrate an HMD-based annotation tool for collecting precise emotion ground truth labels while users are watching 360° videos in Virtual Reality (VR). Our tool uses an HTC VIVE Pro Eye HMD for displaying 360° videos, a Joy-Con controller for inputting emotion annotations, and an Empatica E4 wristband for capturing physiological signals. Timestamps of these devices are synchronized via an NTP server. Following dimensional emotion models, users can report their emotion in terms of valence and arousal as they watch a video in VR. Annotation feedback is provided through two peripheral visualization techniques: HaloLight and DotSize. Our annotation tool provides a starting point for researchers to design momentary and continuous self-reports in virtual environments to enable fine-grained emotion recognition.

With the increasing availability of head-mounted displays (HMDs) that show immersive 360° VR content, it is important to understand to what extent these immersive experiences can evoke emotions. Typically to collect emotion ground truth labels, users rate videos through post-experience self-reports that are discrete in nature. However, post-stimuli self-reports are temporally imprecise, especially after watching 360° videos. In this work, we design six continuous emotion annotation techniques for the Oculus Rift HMD aimed at minimizing workload and distraction. Based on a co-design session with six experts, we contribute HaloLight and DotSize, two continuous annotation methods deemed unobtrusive and easy to understand. We discuss the next challenges for evaluating the usability of these techniques, and reliability of continuous annotations.

Facial thermal imaging has in recent years shown to be an efficient modality for facial emotion recognition. However, the use of deep learning in this field is still not fully exploited given the small number and size of the current datasets. The goal of this work is to improve the performance of the existing deep networks in thermal facial emotion recognition by generating new synthesized thermal images from images in the visual spectrum (RGB). To address this challenging problem, we propose an emotion-guided thermal CycleGAN (ET-CycleGAN). This Generative Adversarial Network (GAN) regularizes the training with facial and emotion priors by extracting features from Convolutional Neural Networks (CNNs) trained for face recognition and facial emotion recognition, respectively. To assess this approach, we generated synthesized images from the training set of the USTC-NVIE dataset, and included the new data to the training set as a data augmentation strategy. By including images generated using the ET-CycleGAN, the accuracy for emotion recognition increased by 10.9%. Our initial findings highlight the importance of adding priors related to training set image attributes (in our case face and emotion priors), to ensure such attributes are maintained in the generated images.

Voice is a rich modality for conveying emotions, however emotional prosody production can be situationally or medically impaired. Since thermal displays have been shown to evoke emotions, we explore how thermal stimulation can augment perception of neutrally-spoken voice messages with affect. We designed ThermalWear, a wearable on-chest thermal display, then tested in a controlled study (N=12) the effects of fabric, thermal intensity, and direction of change. Thereafter, we synthesized 12 neutrally-spoken voice messages, validated (N=7) them, then tested (N=12) if thermal stimuli can augment their perception with affect. We found warm and cool stimuli (a) can be perceived on the chest, and quickly without fabric (4.7-5s) (b) do not incur discomfort (c) generally increase arousal of voice messages and (d) increase / decrease message valence, respectively. We discuss how thermal displays can augment voice perception, which can enhance voice assistants and support individuals with emotional prosody impairments.

Collecting accurate and precise emotion ground truth labels for mobile video watching is essential for ensuring meaningful predictions. However, video-based emotion annotation techniques either rely on post-stimulus discrete self-reports, or allow real-time, continuous emotion annotations (RCEA) only for desktop settings. Following a user-centric approach, we designed an RCEA technique for mobile video watching, and validated its usability and reliability in a controlled, indoor (N=12) and later outdoor (N=20) study. Drawing on physiological measures, interaction logs, and subjective workload reports, we show that (1) RCEA is perceived to be usable for annotating emotions while mobile video watching, without increasing users' mental workload (2) the resulting time-variant annotations are comparable with intended emotion attributes of the video stimuli (classification error for valence: 8.3%; arousal: 25%). We contribute a validated annotation technique and associated annotation fusion method, that is suitable for collecting fine-grained emotion annotations while users watch mobile videos.

To recognize emotions using less obtrusive wearable sensors, we present a novel emotion recognition method that uses only pupil diameter (PD) and skin conductance (SC). Psychological studies show that these two signals are related to the attention level of humans exposed to visual stimuli. Based on this, we propose a feature extraction algorithm that extract correlation-based features for participants watching the same video clip. To boost performance given limited data, we implement a learning system without a deep architecture to classify arousal and valence. Our method outperforms not only state-of-art approaches, but also widely-used traditional and deep learning methods.

Millions of photos are shared online daily, but the richness of interaction compared with face-to-face (F2F) sharing is still missing. While this may change with social Virtual Reality (socialVR), we still lack tools to measure such immersive and interactive experiences. In this paper, we investigate photo sharing experiences in immersive environments, focusing on socialVR. Running context mapping (N=10), an expert creative session (N=6), and an online experience clustering questionnaire (N=20), we develop and statistically evaluate a questionnaire to measure photo sharing experiences. We then ran a controlled, within-subject study (N=26 pairs) to compare photo sharing under F2F, Skype, and Facebook Spaces. Using interviews, audio analysis, and our questionnaire, we found that socialVR can closely approximate F2F sharing. We contribute empirical findings on the immersiveness differences between digital communication media, and propose a socialVR questionnaire that can in the future generalize beyond photo sharing.