Augmented Reality Head-Up Displays (AR-HUDs) have emerged as a transformative technology in the automotive sector, significantly enhancing driver awareness and safety by seamlessly integrating critical information into the driver’s direct line of sight. However, implementing effective AR-HUD systems presents several challenges, including designing intuitive yet minimally distracting user interfaces, ensuring accurate spatial registration, and adapting visualizations to different contexts and levels of vehicle automation. This article provides a comprehensive overview of the current literature on automotive AR-HUDs and offers a structured taxonomy that classifies existing studies along two main dimensions: the specific features of the AR-HUD interface, which capture the relationship between display-related and interaction-related characteristics across tasks and driving contexts, and the human-automation roles, which frame the evolving shift of control, attention, and responsibility between the human and the automated system as vehicle autonomy increases. Additionally, we critically analyze existing testing methodologies and identify significant gaps, such as the overreliance on testing in virtual environments and the lack of standardized frameworks for the progressive evaluation of AR-HUD interfaces, from conceptual designs to immersive virtual simulations and real-world assessments. We conclude by discussing key open research questions and future research directions needed to overcome current limitations and realize the full potential of AR-HUD technology.

Vision-Language Pretrained Models (VLPs) have shown remarkable success in transferring knowledge to various downstream tasks, ranging from image-level tasks such as classification to pixel-level tasks such as Semantic Segmentation. However, a persistent challenge in the latter dense prediction tasks is the misalignment between pixel and text features. This mismatch hinders effective fusion between visual and textual representations, and leads to suboptimal predictions. While some studies attribute this to a Modality Gap, where vision and language modalities form distinct clusters within the shared feature space, we argue that the key issue is semantic misalignment, where the pixel features do not accurately reflect the concepts encoded by the text features. To achieve a stronger semantic alignment between pixels and text embeddings, in this work we propose a Mask-Text Contrastive (MTC) module that explicitly enforces an alignment between image regions and their corresponding semantic concepts. This is achieved by projecting both pixel and text features into a common space where an InfoNCE-based loss promotes semantic correspondence, reducing the modality gap as a side effect. Our approach can be seamlessly integrated into state-of-the-art VLP-based segmentation architectures, requiring only a lightweight linear projection and introducing minimal computational overhead at inference time. Experiments show that the MTC module consistently improves segmentation performance in benchmarks such as ADE20K, COCO-Stuff 10k and Pascal Context. Further experiments with COCO show that MTC is also effective in other downstream dense tasks such as object detection and instance segmentation. The repository associated with this work is available at https://github.com/fedasaro62/mask-text-contrastive-fully-seg.

Cloud gaming represents a rapidly growing segment in the entertainment industry, allowing users to stream and interact with high-quality games over the Internet. However, the problem of maintaining a seamless gaming experience is inherent to minimizing user-perceived latency. In this paper, we present CLoud Application lAtency Prediction (CLAAP), a novel solution that, to tolerate challenged network conditions in gaming, predicts such latency via a Machine Learning (ML) model and forecasts future network evolution. The model, trained over diverse network conditions and gaming scenarios, can then update its parameters via a concept drift detection algorithm that suggests a re-training action, reducing the prediction error up to 21% with minimal overhead. We then integrate this network metrics predictor into a game state prediction to further tolerate network latency spikes even from the user perspective, who can continue playing even in adversarial conditions without session interruptions. The results suggest the potential of advanced predictive analytics in mitigating latency issues, thereby setting the stage for more responsive and immersive cloud gaming services.

Human activities exhibit a strong correlation between actions and the places where these are performed, such as washing something at a sink. More specifically, in daily living environments we may identify particular locations, hereinafter named activity-centric zones , which may afford a set of homogeneous actions. Their knowledge can serve as a prior to favor vision models to recognize human activities. However, the appearance of these zones is scene-specific, limiting the transferability of this prior information to unfamiliar areas and domains. This problem is particularly relevant in egocentric vision, where the environment takes up most of the image, making it even more difficult to separate the action from the context. In this paper, we discuss the importance of decoupling the domain-specific appearance of activity-centric zones from their universal, domain-agnostic representations, and show how the latter can improve the cross-domain transferability of Egocentric Action Recognition (EAR) models. We validate our solution on the EPIC-Kitchens-100 and Argo1M datasets. • We shed light on the side-effects of co-occurrence biases in egocentric vision. • We present EgoZAR, an action recognition model leveraging environmental affordances. • We achieve SOTA performance on EPIC-Kitchens-100 in the domain generalization setting.

Abstract Multi-modal learning has gained significant attention due to its ability to enhance machine learning algorithms. However, it brings challenges related to modality heterogeneity and domain shift. In this work, we address these challenges by proposing a new approach called Relative Norm Alignment (RNA) loss. RNA loss exploits the observation that variations in marginal distributions between modalities manifest as discrepancies in their mean feature norms, and rebalances feature norms across domains, modalities, and classes. This rebalancing improves the accuracy of models on test data from unseen (“target”) distributions. In the context of Unsupervised Domain Adaptation (UDA), we use unlabeled target data to enhance feature transferability. We achieve this by combining RNA loss with an adversarial domain loss and an Information Maximization term that regularizes predictions on target data. We present a comprehensive analysis and ablation of our method for both Domain Generalization and UDA settings, testing our approach on different modalities for tasks such as first and third person action recognition, object recognition, and fatigue detection. Experimental results show that our approach achieves competitive or state-of-the-art performance on the proposed benchmarks, showing the versatility and effectiveness of our method in a wide range of applications.

This paper presents an analysis of the KPT system for the 2022 NIST Language Recognition Evaluation. The KPT submission focuses on the fixed training condition where only specific speech data can be used to develop all the modules and auxiliary systems used to build the language recognizer. Our solution consists of several sub-systems based on different neural network front-ends and a common back-end for classification and fusion. The goal of each front-end is to extract language-related embeddings. Gaussian linear models are used to classify the embeddings of each front-end, followed by multi-class logistic regression to calibrate and fuse the different sub-systems. Experimental results from the NIST LRE 2022 evaluation task show that our approach achieves competitive performance.

Upper Limb (UL) Rehabilitation in Multiple Scle- rosis (MS) is an open research field due to the complex interplay between cognitive and physical dysfunctions. Virtual Reality (VR) can face such an issue by enriching physical training with engaging features, including biofeedback strategies to self- regulate autonomic functions according to the visualisation of indices like heart rate variability (HRV). In the present work, HRV biofeedback is introduced in a VR-based exergame (a game designed to promote exercising), tailored to rehabilitation of the dominant upper limb in Persons with MS (PwMS). The exergame is based on a dual-task paradigm, integrating a UL motor rehabilitative task with a breathing task. The aim is to investigate how the design developed for the HRV biofeedback affects engagement and performance during the exergame session. As a preliminary study, sixteen able-bodied subjects are tested in a within-subjects design, to assess the quality of the game features and design, before approaching MS patients. Two conditions are presented, with and without biofeedback. The proposed HRV biofeedback has two possible levels, depending on whether or not the desired respiratory rate of six breaths/min is successfully maintained. It is used to control game elements and change difficulty of the session. The main finding of this study is that biofeedback improves both user performance and experience in healthy subjects. These results underline the great potential of this technique to promote engagement. Thus, they point to fostering the rehabilitative effectiveness of repetitive motor tasks and encouraging adherence to the long- term training. Future studies will encompass fine tuning of the experimental setup and include PwMS to further adjust the game to patients' needs and observe the setup compliance to rehabilitation settings.

In automated Visual GUI Testing (VGT) for Android devices, the available tools often suffer from low robustness to mobile fragmentation, leading to incorrect results when running the same tests on different devices. To soften these issues, we evaluate two feature matching-based approaches for widget detection in VGT scripts, which use, respectively, the complete full-screen snapshot of the application ( Fullscreen ) and the cropped images of its widgets ( Cropped ) as visual locators to match on emulated devices. Our analysis includes validating the portability of different feature-based visual locators over various apps and devices and evaluating their robustness in terms of cross-device portability and correctly executed interactions. We assessed our results through a comparison with two state-of-the-art tools, EyeAutomate and Sikuli. Despite a limited increase in the computational burden, our Fullscreen approach outperformed state-of-the-art tools in terms of correctly identified locators across a wide range of devices and led to a 30% increase in passing tests. Our work shows that VGT tools’ dependability can be improved by bridging the testing and computer vision communities. This connection enables the design of algorithms targeted to domain-specific needs and thus inherently more usable and robust.

Wearable cameras are becoming more and more popular in several applications, increasing the interest of the research community in developing approaches for recognizing actions from the first-person point of view. An open challenge in egocentric action recognition is that videos lack detailed information about the main actor's pose and thus tend to record only parts of the movement when focusing on manipulation tasks. Thus, the amount of information about the action itself is limited, making crucial the understanding of the manipulated objects and their context. Many previous works addressed this issue with two-stream architectures, where one stream is dedicated to modeling the appearance of objects involved in the action, and another to extracting motion features from optical flow. In this paper, we argue that learning features jointly from these two information channels is beneficial to capture the spatio-temporal correlations between the two better. To this end, we propose a single stream architecture able to do so, thanks to the addition of a self-supervised block that uses a pretext motion prediction task to intertwine motion and appearance knowledge. Experiments on several publicly available databases show the power of our approach.

Biometric identification systems based on fingerprints are vulnerable to attacks that use fake replicas of real fingerprints. One possible countermeasure to this issue consists in developing software modules capable of telling the liveness of an input image and, thus, of discarding fakes prior to the recognition step. This paper presents a fingerprint liveness detection method founded on a patch-based voting approach. Fingerprint images are first segmented to discard background information. Then, small-sized foreground patches are extracted and processed by a well-know Convolutional Neural Network model adapted to the problem at hand. Finally, the patch scores are combined to draw the final fingerprint label. Experimental results on well-established benchmarks demonstrate a promising performance of the proposed method compared with several state-of-the-art algorithms.

Automatic Kinship verification aims at recognizing the degree of kinship of two individuals from their facial images and it has possible applications in image retrieval and annotation, forensics and historical studies. This is a recent and challenging problem, which must deal with different degrees of kinship and variations in age and gender. Our work explores the computer identification of parent–child pairs using a combination of (i) features of different natures, based on geometric and textural data, (ii) feature selection and (iii) state-of-the-art classifiers. Experiments show that the proposed approach provides a valuable solution to the kinship verification problem, as suggested by its comparison with different methods on the same data and the same experimental protocols. We further show the good generalization capabilities of our method in several cross-database experiments.

Kinship verification from facial images in wild conditions is a relatively new and challenging problem in face analysis. Several datasets and algorithms have been proposed in recent years. However, most existing datasets are of small sizes and one standard evaluation protocol is still lack so that it is difficult to compare the performance of different kinship verification methods. In this paper, we present the Kinship Verification in the Wild Competition: the first kinship verification competition which is held in conjunction with the International Joint Conference on Biometrics 2014, Clearwater, Florida, USA. The key goal of this competition is to compare the performance of different methods on a new-collected dataset with the same evaluation protocol and develop the first standardized benchmark for kinship verification in the wild.

Locating sensors in 2D can be often modelled as an Art Gallery problem. Tasks such as surveillance require observing or "covering" the interior of a polygon with a minimum number of sensors or "guards". Observing the boundaries of a polygonal environment is sufficient for tasks such as inspection and image based rendering. As interior covering, also Edge Covering (EC) is NP-hard, and no finite algorithm is known for its exact solution. A number of heuristics have been proposed for the approximate solution of this important problem, but their performances with respect to optimality is unknown. Therefore, a polygon specific tight lower bound for the number of sensors is very useful for assessing the performances of these algorithms. In this paper, we propose a new lower bound for the EC problem. It can be computed in reasonable time for environments with up to a few hundreds of edges. To evaluate its closeness to optimality, we compare it with a previously developed lower bound and with the solution provided by a recent incremental EC algorithm. Tests over hundreds of polygons with different number of edges show that the new lower bound is tight and outperforms the previous one.

Several problems in computer vision require locating multiple sensors. In some cases the problem, which is in general three-dimensional, can be reduced to 2D. This problem can be modeled as an art gallery problem, which is NP-hard and no finite algorithm, even exponential, is known for its solution. Algorithms able to closely approximate the optimal solution and computationally feasible in the worst case are unlikely to exist. In this paper we propose a new sensor locating incremental algorithm. The technique converges toward the optimal solution. It locally refines a starting approximation provided by an integer covering algorithm, where each edge is observed entirely by at least one sensor. A lower bound for the number of sensors, specific of the polygon considered, is used for halting the algorithm, and a set of rules that allow to simplify the problem are presented

The visual hull summarizes the relations between an object and its silhouettes and shadows. This paper develops the theory of the visual hull of general piecewise smooth objects, as those used in CAD applications. A complete catalogue of the nine types of ruled surfaces that are possible boundaries of the visual hull of these objects is derived. The construction of the visual hull is simplified by a detailed analysis that allows pruning and trimming many surfaces not relevant for a particular object. An algorithm for computing the visual hull is presented, together with several examples constructed with a commercial CAD package. The theory developed includes as particular cases the previous approaches to the computation of the visual hull.

The visual hull is a geometric entity that relates the shape of an object to its silhouettes or shadows. This paper develops the theory of the visual hull of generic smooth objects. We show that the visual hull can be constructed using surfaces which partition the viewpoint space of the aspect graph of the object. The surfaces are those generated by the visual events tangent crossing and triple point. An analysis based on the shape of the object at the tangency points of these surfaces allows pruning away many surfaces and patches not relevant to the construction. An algorithm for computing the visual hull is outlined.

3D shapes can be reconstructed from 2D silhouettes by back-projecting them from the corresponding viewpoints and intersecting the resulting solid cones. However, in many practical cases as observing an aircraft or an asteroid, the positions of the viewpoints with respect to the object are not known. In these cases, the relative position of the solid cones is not known and the intersection cannot be performed. The purpose of this paper is introducing and stating in a theoretical framework the problem of understanding 3D shapes from silhouettes when the relative positions of the viewpoints are unknown. The results presented provide a first insight into the problem. In particular, the case of orthographic viewing directions parallel to the same plane is thoroughly discussed, and sets of inequalities are presented which allow determining objects compatible with the silhouettes.