In dyadic interactions, a broad spectrum of human facial reactions might be appropriate for responding to each human speaker behaviour. Following the successful organisation of the REACT 2023 and REACT 2024 challenges, we are proposing the REACT 2025 challenge encouraging the development and benchmarking of Machine Learning (ML) models that can be used to generate multiple appropriate, diverse, realistic and synchronised human-style facial reactions expressed by human listeners in response to an input stimulus (i.e., audio-visual behaviours expressed by their corresponding speakers). As a key of the challenge, we provide challenge participants with the first natural and large-scale multi-modal Multiple Appropriate Facial Reaction Generation (MAFRG) dataset (called MARS) recording 136 human-human dyadic interactions containing a total of 2856 interaction sessions covering five different topics. In addition, this paper also presents the challenge guidelines and the performance of our baselines on the two proposed sub-challenges: Offline MAFRG and Online MAFRG, respectively. The challenge baseline code is publicly available at https://github.com/reactmultimodalchallenge/baseline_react2025

We present a multi-modal stress dataset that uses digital job interviews to induce stress. The dataset provides multi-modal data of 40 participants including audio, video (motion capturing, facial landmarks, eye tracking), as well as physiological information (photoplethysmography, electrodermal activity). In addition to that, the dataset contains time-continuous annotations for stress and occurred emotions (e.g., shame, anger, anxiety, and surprise). In order to establish a baseline, five different machine learning classifiers (Support Vector Machine, K-Nearest Neighbors, Random Forest, Feed-forward Neural Network, and Long-Short-Term Memory Network) have been trained and evaluated on the presented dataset for a binary stress classification task. The best-performing classifier has been a Long-Short-Term Memory Network, which achieved an accuracy of 91.7% and an F1-score of 90.2%. The ForDigitStress dataset is freely available to other researchers.

In dyadic interactions, humans communicate their intentions and state of mind using verbal and non-verbal cues, where multiple different facial reactions might be appropriate in response to a specific speaker behaviour. Then, how to develop a machine learning (ML) model that can automatically generate multiple appropriate, diverse, realistic and synchronised human facial reactions from an previously unseen speaker behaviour is a challenging task. Following the successful organisation of the first REACT challenge (REACT 2023), this edition of the challenge (REACT 2024) employs a subset used by the previous challenge, which contains segmented 30-secs dyadic interaction clips originally recorded as part of the NOXI and RECOLA datasets, encouraging participants to develop and benchmark Machine Learning (ML) models that can generate multiple appropriate facial reactions (including facial image sequences and their attributes) given an input conversational partner's stimulus under various dyadic video conference scenarios. This paper presents: (i) the guidelines of the REACT 2024 challenge; (ii) the dataset utilized in the challenge; and (iii) the performance of the baseline systems on the two proposed sub-challenges: Offline Multiple Appropriate Facial Reaction Generation and Online Multiple Appropriate Facial Reaction Generation, respectively. The challenge baseline code is publicly available at https://github.com/reactmultimodalchallenge/baseline_react2024.

Caregiver mental health disorders increase the risk of insecure infant attachment and can negatively impact multiple aspects of child development, including cognitive, emotional, and social growth. Infant-caregiver interactions contain subtle psychological and behavioral cues that reveal these adverse effects, underscoring the need for analytical methods to assess them effectively. The Face-to-Face-Still-Face (FFSF) paradigm is a key approach in psychological research for investigating these dynamics, and the Infant and Caregiver Engagement Phases revised German edition (ICEP-R) annotation scheme provides a structured framework for evaluating FFSF interactions. However, manual annotation is labor-intensive and limits scalability, thus hindering a deeper understanding of early developmental impairments. To address this, we developed a computational method that automates the annotation of caregiver-infant interactions using features extracted from audio-visual foundational models. Our approach was tested on 92 FFSF video sessions. Findings demonstrate that models based on bidirectional LSTM and linear classifiers show varying effectiveness depending on the role and feature modality. Specifically, bidirectional LSTM models generally perform better in predicting complex infant engagement phases across multimodal features, while linear models show competitive performance, particularly with unimodal feature encodings like Wav2Vec2-BERT. To support further research, we share our raw feature dataset annotated with ICEP-R labels, enabling broader refinement of computational methods in this area.

Estimating the momentary level of participant's engagement is an important prerequisite for assistive systems that support human interactions. Previous work has addressed this task in within-domain evaluation scenarios, i.e. training and testing on the same dataset. This is in contrast to real-life scenarios where domain shifts between training and testing data frequently occur. With MultiMediate'24, we present the first challenge addressing multi-domain engagement estimation. As training data, we utilise the NOXI database of dyadic novice-expert interactions. In addition to within-domain test data, we add two new test domains. First, we introduce recordings following the NOXI protocol but covering languages that are not present in the NOXI training data. Second, we collected novel engagement annotations on the MPIIGroupInteraction dataset which consists of group discussions between three to four people. In this way, MultiMediate'24 evaluates the ability of approaches to generalise across factors such as language and cultural background, group size, task, and screen-mediated vs. face-to-face interaction. This paper describes the MultiMediate'24 challenge and presents baseline results. In addition, we discuss selected challenge solutions.

Understanding emotions is key to Affective Computing. Emotion recognition focuses on the communicative component of emotions encoded in social signals. This view alone is insufficient for a deeper understanding and computational representation of the internal, subjectively experienced component of emotions. This paper presents a cognition-based method calledDeepas a starting point for deeper computational modeling of the internal component of emotions.Deepincorporates an approach to query individual internal emotional experiences and to represent such information computationally. It combines social signals, verbalized introspection information, context information, and theory-driven knowledge. We apply theDeepmethod to the emotion of shame as an example and compare it to a typical emotion recognition model, highlighting the differences and advantages.

Automatic analysis of human behaviour is a fundamental prerequisite for the creation of machines that can effectively interact with- and support humans in social interactions. In MultiMediate'23, we address two key human social behaviour analysis tasks for the first time in a controlled challenge: engagement estimation and bodily behaviour recognition in social interactions. This paper describes the MultiMediate'23 challenge and presents novel sets of annotations for both tasks. For engagement estimation we collected novel annotations on the NOvice eXpert Interaction (NOXI) database. For bodily behaviour recognition, we annotated test recordings of the MPIIGroupInteraction corpus with the BBSI annotation scheme. In addition, we present baseline results for both challenge tasks.

The Multiple Appropriate Facial Reaction Generation Challenge (REACT2023) is the first competition event focused on evaluating multimedia processing and machine learning techniques for generating human-appropriate facial reactions in various dyadic interaction scenarios, with all participants competing strictly under the same conditions. The goal of the challenge is to provide the first benchmark test set for multi-modal information processing and to foster collaboration among the audio, visual, and audio-visual behaviour analysis and behaviour generation (a.k.a generative AI) communities, to compare the relative merits of the approaches to automatic appropriate facial reaction generation under different spontaneous dyadic interaction conditions. This paper presents: (i) the novelties, contributions and guidelines of the REACT2023 challenge; (ii) the dataset utilized in the challenge; and (iii) the performance of the baseline systems on the two proposed sub-challenges: Offline Multiple Appropriate Facial Reaction Generation and Online Multiple Appropriate Facial Reaction Generation, respectively. The challenge baseline code is publicly available at https://github.com/reactmultimodalchallenge/baseline_react2023.

In this article, we introduce a next-generation annotation tool calledNOVAfor emotional behaviour analysis, which implements a workflow that interactively incorporates the ‘human in the loop’. A main aspect of NOVA is the possibility of applying semi-supervised active learning where Machine Learning techniques are used already during the annotation process by giving the possibility to pre-label data automatically. Furthermore, NOVA implements recent eXplainable AI (XAI) techniques to provide users with both, a confidence value of the automatically predicted annotations, as well as visual explanations. We investigate how such techniques can assist non-experts in terms of trust, perceived self-efficacy, cognitive workload as well as creating correct mental models about the system by conducting a user study with 53 participants. The results show that NOVA can easily be used by non-experts and lead to a high computer self-efficacy. Furthermore, the results indicate that XAI visualisations help users to create more correct mental models about the machine learning system compared to the baseline condition. Nevertheless, we suggest that explanations in the field of AI have to be more focused on user-needs as well as on the classification task and the model they want to explain.

Understanding emotions is key to Affective Computing. Emotion recognition focuses on the communicative component of emotions encoded in social signals. This view alone is insufficient for deeper understanding and computational representation of the internal, subjectively experienced component of emotions. This paper presents the Deep method as a starting point for a deeper computational modeling of internal emotions. The method includes how to query individual internal emotional experiences, and it shows an approach to represent such information computationally. It combines social signals, verbalized introspection information, context information, and theory-driven knowledge. We apply the Deep method exemplary on the emotion shame and present a schematic dynamic Bayesian network for modeling it.

In this paper, we introduce a next-generation annotation tool called NOVA, which implements a workflow that interactively incorporates the `human in the loop'. In particular, NOVA offers a collaborative annotation backend where multiple annotators join their workforce. A main aspect of NOVA is the possibility of applying semi-supervised active learning where Machine Learning techniques are used already during the annotation process by giving the possibility to pre-label data automatically. Furthermore, NOVA implements recent eXplainable AI (XAI) techniques to provide users with both, a confidence value of the automatically predicted annotations, as well as visual explanation. This way, annotators get to understand whether they can trust their ML models, or more annotated data is necessary.

Human interaction partners can deal with interruptions and then resume the interaction. This ability should be emulated by social agents. How fast interruptions are handled might influence the overall impression of an agent. In this paper, we present the results of a user study on how a human dialog partner perceives the be- havior of a virtual agent handling verbal user interruptions with different reaction times. The study goes beyond typical perception experiments by preserving the real-time interaction experience. For the evaluation, we rely on a parametrizable parallelized computa- tional model that represents dialog flow, overlap detection, conflict recognition, and conflict handling in real-time. The evaluation re- sults show that the timing of the agent's interruption handling in interactive human-agent dialogues is related to different interper- sonal attitudes.

In this paper, we focus on experience-based role play with virtual agents to provide young adults at the risk of exclusion with social skill training. We present a scenario-based serious game simulation platform. It comes with a social signal interpretation component, a scripted and autonomous agent dialog and social interaction behavior model, and an engine for 3-D rendering of lifelike virtual social agents in a virtual environment. We show how two training systems developed on the basis of this simulation platform can be used to educate people in showing appropriate socioemotive reactions in job interviews. Furthermore, we give an overview of four conducted studies investigating the effect of the agents' portrayed personality and the appearance of the environment on the players' perception of the characters and the learning experience.

We present a novel multi-lingual database of natural dyadic novice-expert interactions, named NoXi, featuring screen-mediated dyadic human interactions in the context of information exchange and retrieval. NoXi is designed to provide spontaneous interactions with emphasis on adaptive behaviors and unexpected situations (e.g. conversational interruptions). A rich set of audio-visual data, as well as continuous and discrete annotations are publicly available through a web interface. Descriptors include low level social signals (e.g. gestures, smiles), functional descriptors (e.g. turn-taking, dialogue acts) and interaction descriptors (e.g. engagement, interest, and fluidity).

When looking at Socially Interactive Robots, adaptation to the user's preferences plays an important role in today's Human-Robot Interaction to keep interaction interesting and engaging over a long period of time. Findings indicate an increase in user engagement for robots with adaptive behavior and personality, but also that it depends on the task context whether a similar or opposing robot personality is preferred. We present an approach based on Reinforcement Learning, which gets its reward directly from social signals in real-time during the interaction, to quickly learn about and dynamically address individual human preferences. Our scenario involves a Reeti robot in the role of a story teller talking about the main characters in the novel “Alice's Adventures in Wonderland” by generating descriptions with varying degree of introversion/extraversion. After initial simulation results, an interactive prototype is presented which allows to explore the learning process adapting to the human interaction partner's engagement.

In this paper we explore the concept of automatic behavioural feedback loops during social interactions. Behavioural feedback loops (BFL) are rapid processes which analyse the behaviour of the user in realtime and provide the user with live feedback on how to improve the behaviour quality. In this context, we implemented an open source software framework for designing, creating and executing BFL on Android powered mobile devices. To get a better understanding of the effects of BFL on face-to-face social interactions, we conducted a user study and compared between four different BFL types spanning three modalities: tactile, auditory and visual. For the study, the BFL have been designed to improve the users' perception of their speaking time in an effort to create more balanced group discussions. The study yielded valuable insights into the impact of BFL on conversations and how humans react to such systems.

We present a demonstration of the ARIA framework, a modular approach for rapid development of virtual humans for information retrieval that have linguistic, emotional, and social skills and a strong personality. We demonstrate the framework's capabilities in a scenario where `Alice in Wonderland', a popular English literature book, is embodied by a virtual human representing Alice. The user can engage in an information exchange dialogue, where Alice acts as the expert on the book, and the user as an interested novice. Besides speech recognition, sophisticated audio-visual behaviour analysis is used to inform the core agent dialogue module about the user's state and intentions, so that it can go beyond simple chat-bot dialogue. The behaviour generation module features a unique new capability of being able to deal gracefully with interruptions of the agent.

Nonverbal and unconscious behaviour is an important component of daily human-human interaction. This is especially true in situations such as public speaking, job interviews or information sensitive conversations, where researchers have shown that an increased awareness of one's behaviour can improve the outcome of the interaction. With wearable technology, such as Google Glass, we now have the opportunity to augment social interactions and provide realtime feedback on one's behaviour in an unobtrusive way. In this paper we present Logue, a system that provides realtime feedback on the presenters' openness, body energy and speech rate during public speaking. The system analyses the user's nonverbal behaviour using social signal processing techniques and gives visual feedback on a head-mounted display. We conducted two user studies with a staged and a real presentation scenario which yielded that Logue's feedback was perceived helpful and had a positive impact on the speaker's performance.

The outcome of interpersonal interactions depends not only on the contents that we communicate verbally, but also on nonverbal social signals. Because a lack of social skills is a common problem for a significant number of people, serious games and other training environments have recently become the focus of research. In this work, we present NovA ( No n v erbal behavior A nalyzer), a system that analyzes and facilitates the interpretation of social signals automatically in a bidirectional interaction with a conversational agent. It records data of interactions, detects relevant social cues, and creates descriptive statistics for the recorded data with respect to the agent's behavior and the context of the situation. This enhances the possibilities for researchers to automatically label corpora of human--agent interactions and to give users feedback on strengths and weaknesses of their social behavior.

Grounding is essential in human interaction and crucial for social robots collaborating with humans. Gaze plays versatile roles for establishing, maintaining and repairing the common ground. It is combined with parallel modalities and involved in several processes for behavior generation and recognition. We present a uniform modeling approach focusing on the multi-modal, parallel and bidirectional aspects of gaze and their interleaving with the dialog logic.

Grounding is an important process that underlies all human interaction. Hence, it is crucial for building social robots that are expected to collaborate effectively with humans. Gaze behavior plays versatile roles in establishing, maintaining and repairing the common ground. Integrating all these roles in a computational dialog model is a complex task since gaze is generally combined with multiple parallel information modalities and involved in multiple processes for the generation and recognition of behavior. Going beyond related work, we present a modeling approach focusing on these multi-modal, parallel and bi-directional aspects of gaze that need to be considered for grounding and their interleaving with the dialog and task management. We illustrate and discuss the different roles of gaze as well as advantages and drawbacks of our modeling approach based on a first user study with a technically sophisticated shared workspace application with a social humanoid robot.

Non-verbal and unconscious behavior plays an important role for efficient human-to-human communication but are often undervalued when training people to become better communicators. This is particularly true for public speakers who need not only behave according to a social etiquette but do so while generating enthusiasm and interest for dozens if not hundreds of other persons. In this paper we propose the concept of social augmentation using wearable computing with the goal of giving users the ability to continuously monitor their performance as a communicator. To this end we explore interaction modalities and feedback mechanisms which would lend themselves to this task.

Automatic detection and interpretation of social signals carried by voice, gestures, mimics, etc. will play a key-role for next-generation interfaces as it paves the way towards a more intuitive and natural human-computer interaction. The paper at hand introduces Social Signal Interpretation (SSI), a framework for real-time recognition of social signals. SSI supports a large range of sensor devices, filter and feature algorithms, as well as, machine learning and pattern recognition tools. It encourages developers to add new components using SSI's C++ API, but also addresses front end users by offering an XML interface to build pipelines with a text editor. SSI is freely available under GPL at http://openssi.net.