Andreas Bulling

Daniel Habermann, Marvin Schmitt, Lars Kühmichel et al.

Multilevel models (MLMs) are a central building block of the Bayesian workflow. They enable joint, interpretable modeling of data across hierarchical levels and provide a fully probabilistic quantification of uncertainty. Despite their well-recognized advantages, MLMs pose significant computational challenges, often rendering their estimation and evaluation intractable within reasonable time constraints. Recent advances in simulation-based inference offer promising solutions for addressing complex probabilistic models using deep generative networks. However, the utility and reliability of deep learning methods for estimating Bayesian MLMs remains largely unexplored, especially when compared with gold-standard samplers. To this end, we explore a family of neural network architectures that leverage the probabilistic factorization of multilevel models to facilitate efficient neural network training and subsequent near-instant posterior inference on unseen datasets. We test our method on several real-world case studies and provide comprehensive comparisons to Stan's gold standard sampler, where possible. Finally, we provide an open-source implementation of our methods to stimulate further research in the nascent field of amortized Bayesian inference.

Philipp Müller, Michal Balazia, Tobias Baur et al.

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.

Ahmed Abdou, Ekta Sood, Philipp Müller et al.

Emotional expressions are inherently multimodal -- integrating facial behavior, speech, and gaze -- but their automatic recognition is often limited to a single modality, e.g. speech during a phone call. While previous work proposed crossmodal emotion embeddings to improve monomodal recognition performance, despite its importance, an explicit representation of gaze was not included. We propose a new approach to emotion recognition that incorporates an explicit representation of gaze in a crossmodal emotion embedding framework. We show that our method outperforms the previous state of the art for both audio-only and video-only emotion classification on the popular One-Minute Gradual Emotion Recognition dataset. Furthermore, we report extensive ablation experiments and provide detailed insights into the performance of different state-of-the-art gaze representations and integration strategies. Our results not only underline the importance of gaze for emotion recognition but also demonstrate a practical and highly effective approach to leveraging gaze information for this task.

Adnen Abdessaied, Mihai Bâce, Andreas Bulling

We propose Neuro-Symbolic Visual Dialog (NSVD) -the first method to combine deep learning and symbolic program execution for multi-round visually-grounded reasoning. NSVD significantly outperforms existing purely-connectionist methods on two key challenges inherent to visual dialog: long-distance co-reference resolution as well as vanishing question-answering performance. We demonstrate the latter by proposing a more realistic and stricter evaluation scheme in which we use predicted answers for the full dialog history when calculating accuracy. We describe two variants of our model and show that using this new scheme, our best model achieves an accuracy of 99.72% on CLEVR-Dialog -a relative improvement of more than 10% over the state of the art while only requiring a fraction of training data. Moreover, we demonstrate that our neuro-symbolic models have a higher mean first failure round, are more robust against incomplete dialog histories, and generalise better not only to dialogs that are up to three times longer than those seen during training but also to unseen question types and scenes.

Anna Penzkofer, Simon Schaefer, Florian Strohm et al.

While deep reinforcement learning (RL) agents outperform humans on an increasing number of tasks, training them requires data equivalent to decades of human gameplay. Recent hierarchical RL methods have increased sample efficiency by incorporating information inherent to the structure of the decision problem but at the cost of having to discover or use human-annotated sub-goals that guide the learning process. We show that intentions of human players, i.e. the precursor of goal-oriented decisions, can be robustly predicted from eye gaze even for the long-horizon sparse rewards task of Montezuma's Revenge - one of the most challenging RL tasks in the Atari2600 game suite. We propose Int-HRL: Hierarchical RL with intention-based sub-goals that are inferred from human eye gaze. Our novel sub-goal extraction pipeline is fully automatic and replaces the need for manual sub-goal annotation by human experts. Our evaluations show that replacing hand-crafted sub-goals with automatically extracted intentions leads to a HRL agent that is significantly more sample efficient than previous methods.

Mayar Elfares, Pascal Reisert, Ralf Küsters et al.

Gaze-based applications are increasingly advancing with the availability of large datasets but ensuring data quality presents a substantial challenge when collecting data at scale. It further requires different parties to collaborate, therefore, privacy concerns arise. We propose QualitEye--the first method for verifying image-based gaze data quality. QualitEye employs a new semantic representation of eye images that contains the information required for verification while excluding irrelevant information for better domain adaptation. QualitEye covers a public setting where parties can freely exchange data and a privacy-preserving setting where parties cannot reveal their raw data nor derive gaze features/labels of others with adapted private set intersection protocols. We evaluate QualitEye on the MPIIFaceGaze and GazeCapture datasets and achieve a high verification performance (with a small overhead in runtime for privacy-preserving versions). Hence, QualitEye paves the way for new gaze analysis methods at the intersection of machine learning, human-computer interaction, and cryptography.

Zhiming Hu, Zheming Yin, Daniel Haeufle et al.

We present HOIMotion - a novel approach for human motion forecasting during human-object interactions that integrates information about past body poses and egocentric 3D object bounding boxes. Human motion forecasting is important in many augmented reality applications but most existing methods have only used past body poses to predict future motion. HOIMotion first uses an encoder-residual graph convolutional network (GCN) and multi-layer perceptrons to extract features from body poses and egocentric 3D object bounding boxes, respectively. Our method then fuses pose and object features into a novel pose-object graph and uses a residual-decoder GCN to forecast future body motion. We extensively evaluate our method on the Aria digital twin (ADT) and MoGaze datasets and show that HOIMotion consistently outperforms state-of-the-art methods by a large margin of up to 8.7% on ADT and 7.2% on MoGaze in terms of mean per joint position error. Complementing these evaluations, we report a human study (N=20) that shows that the improvements achieved by our method result in forecasted poses being perceived as both more precise and more realistic than those of existing methods. Taken together, these results reveal the significant information content available in egocentric 3D object bounding boxes for human motion forecasting and the effectiveness of our method in exploiting this information.

Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi et al.

We propose MToMnet - a Theory of Mind (ToM) neural network for predicting beliefs and their dynamics during human social interactions from multimodal input. ToM is key for effective nonverbal human communication and collaboration, yet, existing methods for belief modelling have not included explicit ToM modelling or have typically been limited to one or two modalities. MToMnet encodes contextual cues (scene videos and object locations) and integrates them with person-specific cues (human gaze and body language) in a separate MindNet for each person. Inspired by prior research on social cognition and computational ToM, we propose three different MToMnet variants: two involving fusion of latent representations and one involving re-ranking of classification scores. We evaluate our approach on two challenging real-world datasets, one focusing on belief prediction, while the other examining belief dynamics prediction. Our results demonstrate that MToMnet surpasses existing methods by a large margin while at the same time requiring a significantly smaller number of parameters. Taken together, our method opens up a highly promising direction for future work on artificial intelligent systems that can robustly predict human beliefs from their non-verbal behaviour and, as such, more effectively collaborate with humans.

Adnen Abdessaied, Lei Shi, Andreas Bulling

We propose $\mathbb{VD}$-$\mathbb{GR}$ - a novel visual dialog model that combines pre-trained language models (LMs) with graph neural networks (GNNs). Prior works mainly focused on one class of models at the expense of the other, thus missing out on the opportunity of combining their respective benefits. At the core of $\mathbb{VD}$-$\mathbb{GR}$ is a novel integration mechanism that alternates between spatial-temporal multi-modal GNNs and BERT layers, and that covers three distinct contributions: First, we use multi-modal GNNs to process the features of each modality (image, question, and dialog history) and exploit their local structures before performing BERT global attention. Second, we propose hub-nodes that link to all other nodes within one modality graph, allowing the model to propagate information from one GNN (modality) to the other in a cascaded manner. Third, we augment the BERT hidden states with fine-grained multi-modal GNN features before passing them to the next $\mathbb{VD}$-$\mathbb{GR}$ layer. Evaluations on VisDial v1.0, VisDial v0.9, VisDialConv, and VisPro show that $\mathbb{VD}$-$\mathbb{GR}$ achieves new state-of-the-art results across all four datasets.

Adnen Abdessaied, Lei Shi, Andreas Bulling

We present MST-MIXER - a novel video dialog model operating over a generic multi-modal state tracking scheme. Current models that claim to perform multi-modal state tracking fall short of two major aspects: (1) They either track only one modality (mostly the visual input) or (2) they target synthetic datasets that do not reflect the complexity of real-world in the wild scenarios. Our model addresses these two limitations in an attempt to close this crucial research gap. Specifically, MST-MIXER first tracks the most important constituents of each input modality. Then, it predicts the missing underlying structure of the selected constituents of each modality by learning local latent graphs using a novel multi-modal graph structure learning method. Subsequently, the learned local graphs and features are parsed together to form a global graph operating on the mix of all modalities which further refines its structure and node embeddings. Finally, the fine-grained graph node features are used to enhance the hidden states of the backbone Vision-Language Model (VLM). MST-MIXER achieves new state-of-the-art results on five challenging benchmarks.

Mayar Elfares, Pascal Reisert, Tilman Dietz et al.

Large language models (LLMs) excel at many general-purpose natural language processing tasks. However, their ability to perform deep reasoning and mathematical analysis, particularly for complex tasks as required in cryptography, remains poorly understood, largely due to the lack of suitable data for evaluation and training. To address this gap, we present CryptoQA, the first large-scale question-answering (QA) dataset specifically designed for cryptography. CryptoQA contains over two million QA pairs drawn from curated academic sources, along with contextual metadata that can be used to test the cryptographic capabilities of LLMs and to train new LLMs on cryptographic tasks. We benchmark 15 state-of-the-art LLMs on CryptoQA, evaluating their factual accuracy, mathematical reasoning, consistency, referencing, backward reasoning, and robustness to adversarial samples. In addition to quantitative metrics, we provide expert reviews that qualitatively assess model outputs and establish a gold-standard baseline. Our results reveal significant performance deficits of LLMs, particularly on tasks that require formal reasoning and precise mathematical knowledge. This shows the urgent need for LLM assistants tailored to cryptography research and development. We demonstrate that, by using CryptoQA, LLMs can be fine-tuned to exhibit better performance on cryptographic tasks.

Constantin Ruhdorfer, Matteo Bortoletto, Anna Penzkofer et al.

We introduce the Overcooked Generalisation Challenge (OGC) - a new benchmark for evaluating reinforcement learning (RL) agents on their ability to cooperate with unknown partners in unfamiliar environments. Existing work typically evaluated cooperative RL only in their training environment or with their training partners, thus seriously limiting our ability to understand agents' generalisation capacity - an essential requirement for future collaboration with humans. The OGC extends Overcooked-AI to support dual curriculum design (DCD). It is fully GPU-accelerated, open-source, and integrated into the minimax DCD benchmark suite. Compared to prior DCD benchmarks, where designers manipulate only minimal elements of the environment, OGC introduces a significantly richer design space: full kitchen layouts with multiple objects that require the designer to account for interaction dynamics between agents. We evaluate state-of-the-art DCD algorithms alongside scalable neural architectures and find that current methods fail to produce agents that generalise effectively to novel layouts and unfamiliar partners. Our results indicate that both agents and curriculum designers struggle with the joint challenge of partner and environment generalisation. These findings establish OGC as a demanding testbed for cooperative generalisation and highlight key directions for future research. We open-source our code.

Moritz Kassner, William Patera, Andreas Bulling

Commercial head-mounted eye trackers provide useful features to customers in industry and research but are expensive and rely on closed source hardware and software. This limits the application areas and use of mobile eye tracking to expert users and inhibits user-driven development, customisation, and extension. In this paper we present Pupil -- an accessible, affordable, and extensible open source platform for mobile eye tracking and gaze-based interaction. Pupil comprises 1) a light-weight headset with high-resolution cameras, 2) an open source software framework for mobile eye tracking, as well as 3) a graphical user interface (GUI) to playback and visualize video and gaze data. Pupil features high-resolution scene and eye cameras for monocular and binocular gaze estimation. The software and GUI are platform-independent and include state-of-the-art algorithms for real-time pupil detection and tracking, calibration, and accurate gaze estimation. Results of a performance evaluation show that Pupil can provide an average gaze estimation accuracy of 0.6 degree of visual angle (0.08 degree precision) with a latency of the processing pipeline of only 0.045 seconds.

Zhiming Hu, Jiahui Xu, Syn Schmitt et al.

Human eye gaze plays a significant role in many virtual and augmented reality (VR/AR) applications, such as gaze-contingent rendering, gaze-based interaction, or eye-based activity recognition. However, prior works on gaze analysis and prediction have only explored eye-head coordination and were limited to human-object interactions. We first report a comprehensive analysis of eye-body coordination in various human-object and human-human interaction activities based on four public datasets collected in real-world (MoGaze), VR (ADT), as well as AR (GIMO and EgoBody) environments. We show that in human-object interactions, e.g. pick and place, eye gaze exhibits strong correlations with full-body motion while in human-human interactions, e.g. chat and teach, a person's gaze direction is correlated with the body orientation towards the interaction partner. Informed by these analyses we then present Pose2Gaze, a novel eye-body coordination model that uses a convolutional neural network and a spatio-temporal graph convolutional neural network to extract features from head direction and full-body poses, respectively, and then uses a convolutional neural network to predict eye gaze. We compare our method with state-of-the-art methods that predict eye gaze only from head movements and show that Pose2Gaze outperforms these baselines with an average improvement of 24.0% on MoGaze, 10.1% on ADT, 21.3% on GIMO, and 28.6% on EgoBody in mean angular error, respectively. We also show that our method significantly outperforms prior methods in the sample downstream task of eye-based activity recognition. These results underline the significant information content available in eye-body coordination during daily activities and open up a new direction for gaze prediction.

Matteo Bortoletto, Constantin Ruhdorfer, Adnen Abdessaied et al.

Recent work on dialogue-based collaborative plan acquisition (CPA) has suggested that Theory of Mind (ToM) modelling can improve missing knowledge prediction in settings with asymmetric skill-sets and knowledge. Although ToM was claimed to be important for effective collaboration, its real impact on this novel task remains under-explored. By representing plans as graphs and by exploiting task-specific constraints we show that, as performance on CPA nearly doubles when predicting one's own missing knowledge, the improvements due to ToM modelling diminish. This phenomenon persists even when evaluating existing baseline methods. To better understand the relevance of ToM for CPA, we report a principled performance comparison of models with and without ToM features. Results across different models and ablations consistently suggest that learned ToM features are indeed more likely to reflect latent patterns in the data with no perceivable link to ToM. This finding calls for a deeper understanding of the role of ToM in CPA and beyond, as well as new methods for modelling and evaluating mental states in computational collaborative agents.

Susanne Hindennach, Lei Shi, Filip Miletić et al.

When users perceive AI systems as mindful, independent agents, they hold them responsible instead of the AI experts who created and designed these systems. So far, it has not been studied whether explanations support this shift in responsibility through the use of mind-attributing verbs like "to think". To better understand the prevalence of mind-attributing explanations we analyse AI explanations in 3,533 explainable AI (XAI) research articles from the Semantic Scholar Open Research Corpus (S2ORC). Using methods from semantic shift detection, we identify three dominant types of mind attribution: (1) metaphorical (e.g. "to learn" or "to predict"), (2) awareness (e.g. "to consider"), and (3) agency (e.g. "to make decisions"). We then analyse the impact of mind-attributing explanations on awareness and responsibility in a vignette-based experiment with 199 participants. We find that participants who were given a mind-attributing explanation were more likely to rate the AI system as aware of the harm it caused. Moreover, the mind-attributing explanation had a responsibility-concealing effect: Considering the AI experts' involvement lead to reduced ratings of AI responsibility for participants who were given a non-mind-attributing or no explanation. In contrast, participants who read the mind-attributing explanation still held the AI system responsible despite considering the AI experts' involvement. Taken together, our work underlines the need to carefully phrase explanations about AI systems in scientific writing to reduce mind attribution and clearly communicate human responsibility.

Haodong Yan, Zhiming Hu, Syn Schmitt et al.

Human motion prediction is important for many virtual and augmented reality (VR/AR) applications such as collision avoidance and realistic avatar generation. Existing methods have synthesised body motion only from observed past motion, despite the fact that human eye gaze is known to correlate strongly with body movements and is readily available in recent VR/AR headsets. We present GazeMoDiff - a novel gaze-guided denoising diffusion model to generate stochastic human motions. Our method first uses a gaze encoder and a motion encoder to extract the gaze and motion features respectively, then employs a graph attention network to fuse these features, and finally injects the gaze-motion features into a noise prediction network via a cross-attention mechanism to progressively generate multiple reasonable human motions in the future. Extensive experiments on the MoGaze and GIMO datasets demonstrate that our method outperforms the state-of-the-art methods by a large margin in terms of multi-modal final displacement error (17.3% on MoGaze and 13.3% on GIMO). We further conducted a human study (N=21) and validated that the motions generated by our method were perceived as both more precise and more realistic than those of prior methods. Taken together, these results reveal the significant information content available in eye gaze for stochastic human motion prediction as well as the effectiveness of our method in exploiting this information.

Matteo Bortoletto, Lei Shi, Andreas Bulling

We propose the Intuitive Reasoning Network (IRENE) - a novel neural model for intuitive psychological reasoning about agents' goals, preferences, and actions that can generalise previous experiences to new situations. IRENE combines a graph neural network for learning agent and world state representations with a transformer to encode the task context. When evaluated on the challenging Baby Intuitions Benchmark, IRENE achieves new state-of-the-art performance on three out of its five tasks - with up to 48.9% improvement. In contrast to existing methods, IRENE is able to bind preferences to specific agents, to better distinguish between rational and irrational agents, and to better understand the role of blocking obstacles. We also investigate, for the first time, the influence of the training tasks on test performance. Our analyses demonstrate the effectiveness of IRENE in combining prior knowledge gained during training for unseen evaluation tasks.

Mayar Elfares, Pascal Reisert, Zhiming Hu et al.

Latest gaze estimation methods require large-scale training data but their collection and exchange pose significant privacy risks. We propose PrivatEyes - the first privacy-enhancing training approach for appearance-based gaze estimation based on federated learning (FL) and secure multi-party computation (MPC). PrivatEyes enables training gaze estimators on multiple local datasets across different users and server-based secure aggregation of the individual estimators' updates. PrivatEyes guarantees that individual gaze data remains private even if a majority of the aggregating servers is malicious. We also introduce a new data leakage attack DualView that shows that PrivatEyes limits the leakage of private training data more effectively than previous approaches. Evaluations on the MPIIGaze, MPIIFaceGaze, GazeCapture, and NVGaze datasets further show that the improved privacy does not lead to a lower gaze estimation accuracy or substantially higher computational costs - both of which are on par with its non-secure counterparts.

Lei Shi, Paul Bürkner, Andreas Bulling

We present ActionDiffusion -- a novel diffusion model for procedure planning in instructional videos that is the first to take temporal inter-dependencies between actions into account in a diffusion model for procedure planning. This approach is in stark contrast to existing methods that fail to exploit the rich information content available in the particular order in which actions are performed. Our method unifies the learning of temporal dependencies between actions and denoising of the action plan in the diffusion process by projecting the action information into the noise space. This is achieved 1) by adding action embeddings in the noise masks in the noise-adding phase and 2) by introducing an attention mechanism in the noise prediction network to learn the correlations between different action steps. We report extensive experiments on three instructional video benchmark datasets (CrossTask, Coin, and NIV) and show that our method outperforms previous state-of-the-art methods on all metrics on CrossTask and NIV and all metrics except accuracy on Coin dataset. We show that by adding action embeddings into the noise mask the diffusion model can better learn action temporal dependencies and increase the performances on procedure planning.

Florian Strohm, Mihai Bâce, Andreas Bulling

Reusable embeddings of user behaviour have shown significant performance improvements for the personalised saliency prediction task. However, prior works require explicit user characteristics and preferences as input, which are often difficult to obtain. We present a novel method to extract user embeddings from pairs of natural images and corresponding saliency maps generated from a small amount of user-specific eye tracking data. At the core of our method is a Siamese convolutional neural encoder that learns the user embeddings by contrasting the image and personal saliency map pairs of different users. Evaluations on two public saliency datasets show that the generated embeddings have high discriminative power, are effective at refining universal saliency maps to the individual users, and generalise well across users and images. Finally, based on our model's ability to encode individual user characteristics, our work points towards other applications that can benefit from reusable embeddings of gaze behaviour.

Zhiming Hu, Guanhua Zhang, Zheming Yin et al.

Human hand and head movements are the most pervasive input modalities in extended reality (XR) and are significant for a wide range of applications. However, prior works on hand and head modelling in XR only explored a single modality or focused on specific applications. We present HaHeAE - a novel self-supervised method for learning generalisable joint representations of hand and head movements in XR. At the core of our method is an autoencoder (AE) that uses a graph convolutional network-based semantic encoder and a diffusion-based stochastic encoder to learn the joint semantic and stochastic representations of hand-head movements. It also features a diffusion-based decoder to reconstruct the original signals. Through extensive evaluations on three public XR datasets, we show that our method 1) significantly outperforms commonly used self-supervised methods by up to 74.0% in terms of reconstruction quality and is generalisable across users, activities, and XR environments, 2) enables new applications, including interpretable hand-head cluster identification and variable hand-head movement generation, and 3) can serve as an effective feature extractor for downstream tasks. Together, these results demonstrate the effectiveness of our method and underline the potential of self-supervised methods for jointly modelling hand-head behaviours in extended reality.

Chuhan Jiao, Yao Wang, Guanhua Zhang et al.

We present DiffGaze, a novel method for generating realistic and diverse continuous human gaze sequences on 360° images based on a conditional score-based denoising diffusion model. Generating human gaze on 360° images is important for various human-computer interaction and computer graphics applications, e.g. for creating large-scale eye tracking datasets or for realistic animation of virtual humans. However, existing methods are limited to predicting discrete fixation sequences or aggregated saliency maps, thereby neglecting crucial parts of natural gaze behaviour. Our method uses features extracted from 360° images as condition and uses two transformers to model the temporal and spatial dependencies of continuous human gaze. We evaluate DiffGaze on two 360° image benchmarks for gaze sequence generation as well as scanpath prediction and saliency prediction. Our evaluations show that DiffGaze outperforms state-of-the-art methods on all tasks on both benchmarks. We also report a 21-participant user study showing that our method generates gaze sequences that are indistinguishable from real human sequences.

Zhiming Hu, Daniel Haeufle, Syn Schmitt et al.

We present HOIGaze - a novel learning-based approach for gaze estimation during hand-object interactions (HOI) in extended reality (XR). HOIGaze addresses the challenging HOI setting by building on one key insight: The eye, hand, and head movements are closely coordinated during HOIs and this coordination can be exploited to identify samples that are most useful for gaze estimator training - as such, effectively denoising the training data. This denoising approach is in stark contrast to previous gaze estimation methods that treated all training samples as equal. Specifically, we propose: 1) a novel hierarchical framework that first recognises the hand currently visually attended to and then estimates gaze direction based on the attended hand; 2) a new gaze estimator that uses cross-modal Transformers to fuse head and hand-object features extracted using a convolutional neural network and a spatio-temporal graph convolutional network; and 3) a novel eye-head coordination loss that upgrades training samples belonging to the coordinated eye-head movements. We evaluate HOIGaze on the HOT3D and Aria digital twin (ADT) datasets and show that it significantly outperforms state-of-the-art methods, achieving an average improvement of 15.6% on HOT3D and 6.0% on ADT in mean angular error. To demonstrate the potential of our method, we further report significant performance improvements for the sample downstream task of eye-based activity recognition on ADT. Taken together, our results underline the significant information content available in eye-hand-head coordination and, as such, open up an exciting new direction for learning-based gaze estimation.

Adnen Abdessaied, Manuel von Hochmeister, Andreas Bulling

We present the Object Language Video Transformer (OLViT) - a novel model for video dialog operating over a multi-modal attention-based dialog state tracker. Existing video dialog models struggle with questions requiring both spatial and temporal localization within videos, long-term temporal reasoning, and accurate object tracking across multiple dialog turns. OLViT addresses these challenges by maintaining a global dialog state based on the output of an Object State Tracker (OST) and a Language State Tracker (LST): while the OST attends to the most important objects within the video, the LST keeps track of the most important linguistic co-references to previous dialog turns. In stark contrast to previous works, our approach is generic by nature and is therefore capable of learning continuous multi-modal dialog state representations of the most relevant objects and rounds. As a result, they can be seamlessly integrated into Large Language Models (LLMs) and offer high flexibility in dealing with different datasets and tasks. Evaluations on the challenging DVD (response classification) and SIMMC 2.1 (response generation) datasets show that OLViT achieves new state-of-the-art performance across both datasets.

Matteo Bortoletto, Constantin Ruhdorfer, Andreas Bulling

Most existing Theory of Mind (ToM) benchmarks for foundation models rely on variations of the Sally-Anne test, offering only a very limited perspective on ToM and neglecting the complexity of human social interactions. To address this gap, we propose ToM-SSI: a new benchmark specifically designed to test ToM capabilities in environments rich with social interactions and spatial dynamics. While current ToM benchmarks are limited to text-only or dyadic interactions, ToM-SSI is multimodal and includes group interactions of up to four agents that communicate and move in situated environments. This unique design allows us to study, for the first time, mixed cooperative-obstructive settings and reasoning about multiple agents' mental state in parallel, thus capturing a wider range of social cognition than existing benchmarks. Our evaluations reveal that the current models' performance is still severely limited, especially in these new tasks, highlighting critical gaps for future research.

Constantin Ruhdorfer, Matteo Bortoletto, Victor Oei et al.

We introduce Unsupervised Partner Design (UPD) - a population-free, multi-agent reinforcement learning framework for robust ad-hoc teamwork that adaptively generates training partners without requiring pretrained partners or manual parameter tuning. UPD constructs diverse partners by stochastically mixing an ego agent's policy with biased random behaviours and scores them using a variance-based learnability metric that prioritises partners near the ego agent's current learning frontier. We show that UPD can be integrated with unsupervised environment design, resulting in the first method enabling fully unsupervised curricula over both level and partner distributions in a cooperative setting. Through extensive evaluations on Overcooked-AI and the Overcooked Generalisation Challenge, we demonstrate that this dynamic partner curriculum is highly effective: UPD consistently outperforms both population-based and population-free baselines as well as ablations. In a user study, we further show that UPD achieves higher returns than all baselines and was perceived as significantly more adaptive, more human-like, a better collaborator, and less frustrating.

Adnen Abdessaied, Anna Rohrbach, Marcus Rohrbach et al.

We present V$^2$Dial - a novel expert-based model specifically geared towards simultaneously handling image and video input data for multimodal conversational tasks. Current multimodal models primarily focus on simpler tasks (e.g., VQA, VideoQA, video-text retrieval) and often neglect the more challenging conversational counterparts, such as video and visual/image dialog. Moreover, works on both conversational tasks evolved separately from each other despite their apparent similarities limiting their applicability potential. To this end, we propose to unify both tasks using a single model that for the first time jointly learns the spatial and temporal features of images and videos by routing them through dedicated experts and aligns them using matching and contrastive learning techniques. Furthermore, we systemically study the domain shift between the two tasks by investigating whether and to what extent these seemingly related tasks can mutually benefit from their respective training data. Extensive evaluations on the widely used video and visual dialog datasets of AVSD and VisDial show that our model achieves new state-of-the-art results across four benchmarks both in zero-shot and fine-tuning settings.

Florian Strohm, Mihai Bâce, Andreas Bulling

We present HAIFAI - a novel two-stage system where humans and AI interact to tackle the challenging task of reconstructing a visual representation of a face that exists only in a person's mind. In the first stage, users iteratively rank images our reconstruction system presents based on their resemblance to a mental image. These rankings, in turn, allow the system to extract relevant image features, fuse them into a unified feature vector, and use a generative model to produce an initial reconstruction of the mental image. The second stage leverages an existing face editing method, allowing users to manually refine and further improve this reconstruction using an easy-to-use slider interface for face shape manipulation. To avoid the need for tedious human data collection for training the reconstruction system, we introduce a computational user model of human ranking behaviour. For this, we collected a small face ranking dataset through an online crowd-sourcing study containing data from 275 participants. We evaluate HAIFAI and an ablated version in a 12-participant user study and demonstrate that our approach outperforms the previous state of the art regarding reconstruction quality, usability, perceived workload, and reconstruction speed. We further validate the reconstructions in a subsequent face ranking study with 18 participants and show that HAIFAI achieves a new state-of-the-art identification rate of 60.6%. These findings represent a significant advancement towards developing new interactive intelligent systems capable of reliably and effortlessly reconstructing a user's mental image.

Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi et al.

Despite growing interest in Theory of Mind (ToM) tasks for evaluating language models (LMs), little is known about how LMs internally represent mental states of self and others. Understanding these internal mechanisms is critical - not only to move beyond surface-level performance, but also for model alignment and safety, where subtle misattributions of mental states may go undetected in generated outputs. In this work, we present the first systematic investigation of belief representations in LMs by probing models across different scales, training regimens, and prompts - using control tasks to rule out confounds. Our experiments provide evidence that both model size and fine-tuning substantially improve LMs' internal representations of others' beliefs, which are structured - not mere by-products of spurious correlations - yet brittle to prompt variations. Crucially, we show that these representations can be strengthened: targeted edits to model activations can correct wrong ToM inferences.

Anna Penzkofer, Lei Shi, Andreas Bulling

While Vector Symbolic Architectures (VSAs) are promising for modelling spatial cognition, their application is currently limited to artificially generated images and simple spatial queries. We propose VSA4VQA - a novel 4D implementation of VSAs that implements a mental representation of natural images for the challenging task of Visual Question Answering (VQA). VSA4VQA is the first model to scale a VSA to complex spatial queries. Our method is based on the Semantic Pointer Architecture (SPA) to encode objects in a hyperdimensional vector space. To encode natural images, we extend the SPA to include dimensions for object's width and height in addition to their spatial location. To perform spatial queries we further introduce learned spatial query masks and integrate a pre-trained vision-language model for answering attribute-related questions. We evaluate our method on the GQA benchmark dataset and show that it can effectively encode natural images, achieving competitive performance to state-of-the-art deep learning methods for zero-shot VQA.

Florian Strohm, Mihai Bâce, Andreas Bulling

We present User-predictable Face Editing (UP-FacE) -- a novel method for predictable face shape editing. In stark contrast to existing methods for face editing using trial and error, edits with UP-FacE are predictable by the human user. That is, users can control the desired degree of change precisely and deterministically and know upfront the amount of change required to achieve a certain editing result. Our method leverages facial landmarks to precisely measure facial feature values, facilitating the training of UP-FacE without manually annotated attribute labels. At the core of UP-FacE is a transformer-based network that takes as input a latent vector from a pre-trained generative model and a facial feature embedding, and predicts a suitable manipulation vector. To enable user-predictable editing, a scaling layer adjusts the manipulation vector to achieve the precise desired degree of change. To ensure that the desired feature is manipulated towards the target value without altering uncorrelated features, we further introduce a novel semantic face feature loss. Qualitative and quantitative results demonstrate that UP-FacE enables precise and fine-grained control over 23 face shape features.

Zhiming Hu, Syn Schmitt, Daniel Haeufle et al.

We present GazeMotion, a novel method for human motion forecasting that combines information on past human poses with human eye gaze. Inspired by evidence from behavioural sciences showing that human eye and body movements are closely coordinated, GazeMotion first predicts future eye gaze from past gaze, then fuses predicted future gaze and past poses into a gaze-pose graph, and finally uses a residual graph convolutional network to forecast body motion. We extensively evaluate our method on the MoGaze, ADT, and GIMO benchmark datasets and show that it outperforms state-of-the-art methods by up to 7.4% improvement in mean per joint position error. Using head direction as a proxy to gaze, our method still achieves an average improvement of 5.5%. We finally report an online user study showing that our method also outperforms prior methods in terms of perceived realism. These results show the significant information content available in eye gaze for human motion forecasting as well as the effectiveness of our method in exploiting this information.

Yao Wang, Chuhan Jiao, Mihai Bâce et al.

Despite its importance for assessing the effectiveness of communicating information visually, fine-grained recallability of information visualisations has not been studied quantitatively so far. In this work, we propose a question-answering paradigm to study visualisation recallability and present VisRecall - a novel dataset consisting of 200 visualisations that are annotated with crowd-sourced human (N = 305) recallability scores obtained from 1,000 questions of five question types. Furthermore, we present the first computational method to predict recallability of different visualisation elements, such as the title or specific data values. We report detailed analyses of our method on VisRecall and demonstrate that it outperforms several baselines in overall recallability and FE-, F-, RV-, and U-question recallability. Our work makes fundamental contributions towards a new generation of methods to assist designers in optimising visualisations.

Yao Wang, Mihai Bâce, Andreas Bulling

We propose Unified Model of Saliency and Scanpaths (UMSS) -- a model that learns to predict visual saliency and scanpaths (i.e. sequences of eye fixations) on information visualisations. Although scanpaths provide rich information about the importance of different visualisation elements during the visual exploration process, prior work has been limited to predicting aggregated attention statistics, such as visual saliency. We present in-depth analyses of gaze behaviour for different information visualisation elements (e.g. Title, Label, Data) on the popular MASSVIS dataset. We show that while, overall, gaze patterns are surprisingly consistent across visualisations and viewers, there are also structural differences in gaze dynamics for different elements. Informed by our analyses, UMSS first predicts multi-duration element-level saliency maps, then probabilistically samples scanpaths from them. Extensive experiments on MASSVIS show that our method consistently outperforms state-of-the-art methods with respect to several, widely used scanpath and saliency evaluation metrics. Our method achieves a relative improvement in sequence score of 11.5% for scanpath prediction, and a relative improvement in Pearson correlation coefficient of up to 23.6% for saliency prediction. These results are auspicious and point towards richer user models and simulations of visual attention on visualisations without the need for any eye tracking equipment.

Ekta Sood, Fabian Kögel, Philipp Müller et al.

Human-like attention as a supervisory signal to guide neural attention has shown significant promise but is currently limited to uni-modal integration - even for inherently multimodal tasks such as visual question answering (VQA). We present the Multimodal Human-like Attention Network (MULAN) - the first method for multimodal integration of human-like attention on image and text during training of VQA models. MULAN integrates attention predictions from two state-of-the-art text and image saliency models into neural self-attention layers of a recent transformer-based VQA model. Through evaluations on the challenging VQAv2 dataset, we show that MULAN achieves a new state-of-the-art performance of 73.98% accuracy on test-std and 73.72% on test-dev and, at the same time, has approximately 80% fewer trainable parameters than prior work. Overall, our work underlines the potential of integrating multimodal human-like and neural attention for VQA

Ekta Sood, Fabian Kögel, Florian Strohm et al.

We present VQA-MHUG - a novel 49-participant dataset of multimodal human gaze on both images and questions during visual question answering (VQA) collected using a high-speed eye tracker. We use our dataset to analyze the similarity between human and neural attentive strategies learned by five state-of-the-art VQA models: Modular Co-Attention Network (MCAN) with either grid or region features, Pythia, Bilinear Attention Network (BAN), and the Multimodal Factorized Bilinear Pooling Network (MFB). While prior work has focused on studying the image modality, our analyses show - for the first time - that for all models, higher correlation with human attention on text is a significant predictor of VQA performance. This finding points at a potential for improving VQA performance and, at the same time, calls for further research on neural text attention mechanisms and their integration into architectures for vision and language tasks, including but potentially also beyond VQA.

Florian Strohm, Ekta Sood, Sven Mayer et al.

We propose a novel method that leverages human fixations to visually decode the image a person has in mind into a photofit (facial composite). Our method combines three neural networks: An encoder, a scoring network, and a decoder. The encoder extracts image features and predicts a neural activation map for each face looked at by a human observer. A neural scoring network compares the human and neural attention and predicts a relevance score for each extracted image feature. Finally, image features are aggregated into a single feature vector as a linear combination of all features weighted by relevance which a decoder decodes into the final photofit. We train the neural scoring network on a novel dataset containing gaze data of 19 participants looking at collages of synthetic faces. We show that our method significantly outperforms a mean baseline predictor and report on a human study that shows that we can decode photofits that are visually plausible and close to the observer's mental image.

Ekta Sood, Simon Tannert, Philipp Mueller et al.

A lack of corpora has so far limited advances in integrating human gaze data as a supervisory signal in neural attention mechanisms for natural language processing(NLP). We propose a novel hybrid text saliency model(TSM) that, for the first time, combines a cognitive model of reading with explicit human gaze supervision in a single machine learning framework. On four different corpora we demonstrate that our hybrid TSM duration predictions are highly correlated with human gaze ground truth. We further propose a novel joint modeling approach to integrate TSM predictions into the attention layer of a network designed for a specific upstream NLP task without the need for any task-specific human gaze data. We demonstrate that our joint model outperforms the state of the art in paraphrase generation on the Quora Question Pairs corpus by more than 10% in BLEU-4 and achieves state of the art performance for sentence compression on the challenging Google Sentence Compression corpus. As such, our work introduces a practical approach for bridging between data-driven and cognitive models and demonstrates a new way to integrate human gaze-guided neural attention into NLP tasks.

Ekta Sood, Simon Tannert, Diego Frassinelli et al.

While neural networks with attention mechanisms have achieved superior performance on many natural language processing tasks, it remains unclear to which extent learned attention resembles human visual attention. In this paper, we propose a new method that leverages eye-tracking data to investigate the relationship between human visual attention and neural attention in machine reading comprehension. To this end, we introduce a novel 23 participant eye tracking dataset - MQA-RC, in which participants read movie plots and answered pre-defined questions. We compare state of the art networks based on long short-term memory (LSTM), convolutional neural models (CNN) and XLNet Transformer architectures. We find that higher similarity to human attention and performance significantly correlates to the LSTM and CNN models. However, we show this relationship does not hold true for the XLNet models -- despite the fact that the XLNet performs best on this challenging task. Our results suggest that different architectures seem to learn rather different neural attention strategies and similarity of neural to human attention does not guarantee best performance.

Inken Hagestedt, Michael Backes, Andreas Bulling

An ever-growing body of work has demonstrated the rich information content available in eye movements for user modelling, e.g. for predicting users' activities, cognitive processes, or even personality traits. We show that state-of-the-art classifiers for eye-based user modelling are highly vulnerable to adversarial examples: small artificial perturbations in gaze input that can dramatically change a classifier's predictions. We generate these adversarial examples using the Fast Gradient Sign Method (FGSM) that linearises the gradient to find suitable perturbations. On the sample task of eye-based document type recognition we study the success of different adversarial attack scenarios: with and without knowledge about classifier gradients (white-box vs. black-box) as well as with and without targeting the attack to a specific class, In addition, we demonstrate the feasibility of defending against adversarial attacks by adding adversarial examples to a classifier's training data.

Mihai Bâce, Sander Staal, Andreas Bulling

Quantification of human attention is key to several tasks in mobile human-computer interaction (HCI), such as predicting user interruptibility, estimating noticeability of user interface content, or measuring user engagement. Previous works to study mobile attentive behaviour required special-purpose eye tracking equipment or constrained users' mobility. We propose a novel method to sense and analyse visual attention on mobile devices during everyday interactions. We demonstrate the capabilities of our method on the sample task of eye contact detection that has recently attracted increasing research interest in mobile HCI. Our method builds on a state-of-the-art method for unsupervised eye contact detection and extends it to address challenges specific to mobile interactive scenarios. Through evaluation on two current datasets, we demonstrate significant performance improvements for eye contact detection across mobile devices, users, or environmental conditions. Moreover, we discuss how our method enables the calculation of additional attention metrics that, for the first time, enable researchers from different domains to study and quantify attention allocation during mobile interactions in the wild.

Mihai Bâce, Sander Staal, Andreas Bulling

With an ever-increasing number of mobile devices competing for our attention, quantifying when, how often, or for how long users visually attend to their devices has emerged as a core challenge in mobile human-computer interaction. Encouraged by recent advances in automatic eye contact detection using machine learning and device-integrated cameras, we provide a fundamental investigation into the feasibility of quantifying visual attention during everyday mobile interactions. We identify core challenges and sources of errors associated with sensing attention on mobile devices in the wild, including the impact of face and eye visibility, the importance of robust head pose estimation, and the need for accurate gaze estimation. Based on this analysis, we propose future research directions and discuss how eye contact detection represents the foundation for exciting new applications towards next-generation pervasive attentive user interfaces.

Philipp Müller, Andreas Bulling

Automatic detection of emergent leaders in small groups from nonverbal behaviour is a growing research topic in social signal processing but existing methods were evaluated on single datasets -- an unrealistic assumption for real-world applications in which systems are required to also work in settings unseen at training time. It therefore remains unclear whether current methods for emergent leadership detection generalise to similar but new settings and to which extent. To overcome this limitation, we are the first to study a cross-dataset evaluation setting for the emergent leadership detection task. We provide evaluations for within- and cross-dataset prediction using two current datasets (PAVIS and MPIIGroupInteraction), as well as an investigation on the robustness of commonly used feature channels (visual focus of attention, body pose, facial action units, speaking activity) and online prediction in the cross-dataset setting. Our evaluations show that using pose and eye contact based features, cross-dataset prediction is possible with an accuracy of 0.68, as such providing another important piece of the puzzle towards emergent leadership detection in the real world.

Michael Xuelin Huang, Andreas Bulling

Recent methods to automatically calibrate stationary eye trackers were shown to effectively reduce inherent calibration distortion. However, these methods require additional information, such as mouse clicks or on-screen content. We propose the first method that only requires users' eye movements to reduce calibration distortion in the background while users naturally look at an interface. Our method exploits that calibration distortion makes straight saccade trajectories appear curved between the saccadic start and end points. We show that this curving effect is systematic and the result of distorted gaze projection plane. To mitigate calibration distortion, our method undistorts this plane by straightening saccade trajectories using image warping. We show that this approach improves over the common six-point calibration and is promising for reducing distortion. As such, it provides a non-intrusive solution to alleviating accuracy decrease of eye tracker during long-term use.

Michae Xuelin Huang, Jiajia Li, Grace Ngai et al.

Internal thought refers to the process of directing attention away from a primary visual task to internal cognitive processing. Internal thought is a pervasive mental activity and closely related to primary task performance. As such, automatic detection of internal thought has significant potential for user modelling in intelligent interfaces, particularly for e-learning applications. Despite the close link between the eyes and the human mind, only a few studies have investigated vergence behaviour during internal thought and none has studied moment-to-moment detection of internal thought from gaze. While prior studies relied on long-term data analysis and required a large number of gaze characteristics, we describe a novel method that is computationally light-weight and that only requires eye vergence information that is readily available from binocular eye trackers. We further propose a novel paradigm to obtain ground truth internal thought annotations that exploits human blur perception. We evaluate our method for three increasingly challenging detection tasks: (1) during a controlled math-solving task, (2) during natural viewing of lecture videos, and (3) during daily activities, such as coding, browsing, and reading. Results from these evaluations demonstrate the performance and robustness of vergence-based detection of internal thought and, as such, open up new directions for research on interfaces that adapt to shifts of mental attention.

Seonwook Park, Xucong Zhang, Andreas Bulling et al.

Conventional feature-based and model-based gaze estimation methods have proven to perform well in settings with controlled illumination and specialized cameras. In unconstrained real-world settings, however, such methods are surpassed by recent appearance-based methods due to difficulties in modeling factors such as illumination changes and other visual artifacts. We present a novel learning-based method for eye region landmark localization that enables conventional methods to be competitive to latest appearance-based methods. Despite having been trained exclusively on synthetic data, our method exceeds the state of the art for iris localization and eye shape registration on real-world imagery. We then use the detected landmarks as input to iterative model-fitting and lightweight learning-based gaze estimation methods. Our approach outperforms existing model-fitting and appearance-based methods in the context of person-independent and personalized gaze estimation.

Arif Khan, Ingmar Steiner, Yusuke Sugano et al.

Phonetic segmentation is the process of splitting speech into distinct phonetic units. Human experts routinely perform this task manually by analyzing auditory and visual cues using analysis software, which is an extremely time-consuming process. Methods exist for automatic segmentation, but these are not always accurate enough. In order to improve automatic segmentation, we need to model it as close to the manual segmentation as possible. This corpus is an effort to capture the human segmentation behavior by recording experts performing a segmentation task. We believe that this data will enable us to highlight the important aspects of manual segmentation, which can be used in automatic segmentation to improve its accuracy.

Xucong Zhang, Yusuke Sugano, Mario Fritz et al.

Learning-based methods are believed to work well for unconstrained gaze estimation, i.e. gaze estimation from a monocular RGB camera without assumptions regarding user, environment, or camera. However, current gaze datasets were collected under laboratory conditions and methods were not evaluated across multiple datasets. Our work makes three contributions towards addressing these limitations. First, we present the MPIIGaze that contains 213,659 full face images and corresponding ground-truth gaze positions collected from 15 users during everyday laptop use over several months. An experience sampling approach ensured continuous gaze and head poses and realistic variation in eye appearance and illumination. To facilitate cross-dataset evaluations, 37,667 images were manually annotated with eye corners, mouth corners, and pupil centres. Second, we present an extensive evaluation of state-of-the-art gaze estimation methods on three current datasets, including MPIIGaze. We study key challenges including target gaze range, illumination conditions, and facial appearance variation. We show that image resolution and the use of both eyes affect gaze estimation performance while head pose and pupil centre information are less informative. Finally, we propose GazeNet, the first deep appearance-based gaze estimation method. GazeNet improves the state of the art by 22% percent (from a mean error of 13.9 degrees to 10.8 degrees) for the most challenging cross-dataset evaluation.

Hosnieh Sattar, Mario Fritz, Andreas Bulling

What does human gaze reveal about a users' intents and to which extend can these intents be inferred or even visualized? Gaze was proposed as an implicit source of information to predict the target of visual search and, more recently, to predict the object class and attributes of the search target. In this work, we go one step further and investigate the feasibility of combining recent advances in encoding human gaze information using deep convolutional neural networks with the power of generative image models to visually decode, i.e. create a visual representation of, the search target. Such visual decoding is challenging for two reasons: 1) the search target only resides in the user's mind as a subjective visual pattern, and can most often not even be described verbally by the person, and 2) it is, as of yet, unclear if gaze fixations contain sufficient information for this task at all. We show, for the first time, that visual representations of search targets can indeed be decoded only from human gaze fixations. We propose to first encode fixations into a semantic representation and then decode this representation into an image. We evaluate our method on a recent gaze dataset of 14 participants searching for clothing in image collages and validate the model's predictions using two human studies. Our results show that 62% (Chance level = 10%) of the time users were able to select the categories of the decoded image right. In our second studies we show the importance of a local gaze encoding for decoding visual search targets of user