Realistic and smooth full-body tracking is crucial for immersive AR/VR applications. Existing systems primarily track head and hands via Head Mounted Devices (HMDs) and controllers, making the 3D full-body reconstruction incomplete. One potential approach is to generate the full-body motions from sparse inputs collected from limited sensors using a Neural Network (NN) model. In this paper, we propose a novel method based on a multi-layer perceptron (MLP) backbone that is enhanced with residual connections and a novel NN-component called Memory-Block. In particular, Memory-Block represents missing sensor data with trainable code-vectors, which are combined with the sparse signals from previous time instances to improve the temporal consistency. Furthermore, we formulate our solution as a multi-task learning problem, allowing our MLP-backbone to learn robust representations that boost accuracy. Our experiments show that our method outperforms state-of-the-art baselines by substantially reducing prediction errors. Moreover, it achieves 72 FPS on mobile HMDs that ultimately improves the accuracy-running time tradeoff.

Data augmentation (DA) is ubiquitously used in training of Automatic Speech Recognition (ASR) models. DA offers increased data variability, robustness and generalization against different acoustic distortions. Recently, personalization of ASR models on mobile devices has been shown to improve Word Error Rate (WER). This paper evaluates data augmentation in this context and proposes persoDA; a DA method driven by user’s data utilized to personalize ASR [1] –[3]. persoDA aims to augment training with data specifically tuned towards acoustic characteristics of the end-user, as opposed to standard augmentation based on Multi-Condition Training (MCT) that applies random reverberation and noises. Our evaluation with an ASR conformer-based baseline trained on Librispeech and per-sonalized for VOICES [4] shows that persoDA achieves a 13.9% relative WER reduction over using standard data augmentation (using random noise & reverberation). Furthermore, persoDA shows 16% to 20% faster convergence over MCT.

The proliferation of AI models in everyday devices has highlighted a critical challenge: prediction errors that degrade user experience. While existing solutions focus on error detection, they rarely provide efficient correction mechanisms, especially for resource-constrained devices. We present a novel system enabling users to correct AI misclassifications through few-shot learning, requiring minimal computational resources and storage. Our approach combines server-side foundation model training with on-device prototype-based classification, enabling efficient error correction through prototype updates rather than model retraining. The system consists of two key components: (1) a server-side pipeline that leverages knowledge distillation to transfer robust feature representations from foundation models to device-compatible architectures, and (2) a device-side mechanism that enables ultra-efficient error correction through prototype adaptation. We demonstrate our system's effectiveness on both image classification and object detection tasks, achieving over 50% error correction in one-shot scenarios on Food-101 and Flowers-102 datasets while maintaining minimal forgetting (less than 0.02%) and negligible computational overhead. Our implementation, validated through an Android demonstration app, proves the system's practicality in real-world scenarios.

Air quality intensifies climate change through global pollution, particularly affecting lower and middle-income countries that lack local ground pollutant monitoring networks. While atmospheric pollutant satellite measurements offer a broad view, their coarse spatial resolution limits detailed air pollution insights, resulting in unmonitored regions and information gaps. Furthermore, satellites generate extensive data that is often challenging to directly correlate with ground air pollution stations. To overcome this, we propose a multimodal self-supervised approach that learns from diverse satellite sources for air pollution monitoring. More specifically, aside from incorporating a combination of multi-spectral and spectral modalities (Sentinel-2 & 5P products), we also leverage tabular land cover data. Their integration into multimodal self-supervised learning is highlighted, employing a novel augmentation scheme that results in more resilient embeddings. The proposed approach integrates a self-supervised redundancy reduction loss in a multi-modal fashion, capturing both inter-modal and intra-modal correspondences. Furthermore, an adaptive loss weighting mechanism is introduced to effectively combine different multi-modal losses. Our approach’s efficacy is showcased in the air pollution prediction task, exhibiting a noteworthy improvement of up to 17% compared to existing methods. Furthermore, in our experiments, the applicability of our approach in other environmental tasks is also exhibited.

The primary environmental health threat in the WHO European Region is air pollution, impacting the daily health and well-being of its citizens significantly. To effectively understand the impact, and dynamics of air quality a detailed investigation of different environmental, weather, and land cover indices is appropriate. To this end, this paper introduces three European cities’ spatiotemporal datasets, customized for air pollution monitoring at a regional level. The datasets are composed of major air quality, weather measurements and land use information. The duration is approximately from 2020 to 2023 with an hourly temporal resolution and a spatial resolution of 0.005°. The temporal and spatiotemporal datasets are publicly released aiming to provide a solid foundation for researchers, analysts, and practitioners to conduct in-depth analyses of air pollution dynamics. The ready-to-use data, processing code & examples are available at https://github.com/angelisgiorgos/RegionalAQDatasets

In this study, the critical issue of air pollution and its impact on quality of life is addressed by developing technologies for monitoring air quality and identifying areas of concern. A notable challenge in this domain is the collection of air pollution data, which often lacks differentiation between normal and abnormal quality levels. Recognizing the importance of detecting anomalies in air pollution data, a novel methodology is introduced that not only safeguards human health but also enhances the overall data quality. The proposed approach involves the injection of abnormal events into air pollution datasets by leveraging the temporal distribution characteristics of the data. To assess the validity of these generated anomalies, the Kolmogorov-Smirnoff test is employed. Furthermore, the developed method is evaluated by exploiting SoA deep learning based models for anomalies detection. Finally, a deep attention-based autencoder method is designed, namely AT-MCRAAD that demonstrates superior performance over existing traditional and contemporary algorithms in identifying these anomalous events.

Neural Radiance Fields (NeRF) have quickly become the primary approach for 3D reconstruction and novel view synthesis in recent years due to their remarkable performance. Despite the huge interest in NeRF methods, a practical use case of NeRFs has largely been ignored; the exploration of the scene space modelled by a NeRF. In this paper, for the first time in the literature, we propose and formally define the scene exploration framework as the efficient discovery of NeRF model inputs (i.e. coordinates and viewing angles), using which one can render novel views that adhere to user-selected criteria. To remedy the lack of approaches addressing scene exploration, we first propose two baseline methods called Guided-Random Search (GRS) and Pose Interpolation-based Search (PIBS). We then cast scene exploration as an optimization problem, and propose the criteria-agnostic Evolution-Guided Pose Search (EGPS) for efficient exploration. We test all three approaches with various criteria (e.g. saliency maximization, image quality maximization, photo-composition quality improvement) and show that our EGPS performs more favourably than other baselines. We finally highlight key points and limitations, and outline directions for future research in scene exploration.

Novel networking paradigms are enabling the introduction of innovative vertical use cases and new business relations in the B5G/6G mobile ecosystem. For instance, a use case may require the coordination of domains owned by different operators to provide service continuity and keep offering a vertical network service (NS) in similar conditions after a cross-border situation. This demonstration presents a procedure to perform a proactive service relocation in such a multi-administrative domain scenario considering an automotive use case. This demonstration proposes a cloud-native solution combining multiple enablers to manage the life-cycle of virtualised automotive NSs. During run-time, and upon registration, an AI/ML-based enabler decides proactively on the service relocation moment based on the collected vehicle’s positions and triggers an Integration Fabric enabler following ETSI ZSM guidelines to start a new instance of such automotive NS at the associated ETSI NFV management and orchestration stack present in the neighbouring administrative domain.

Automatic Speech Recognition (ASR) still face challenges when recognizing time-variant rare-phrases. Contextual biasing (CB) modules bias ASR model towards such contextually-relevant phrases. During training, a list of biasing phrases are selected from a large pool of phrases following a sampling strategy. In this work we firstly analyse different sampling strategies to provide insights into the training of CB for ASR with correlation plots between the bias embeddings among various training stages. Secondly, we introduce a neighbourhood attention (NA) that localizes self attention (SA) to the nearest neighbouring frames to further refine the CB output. The results show that this proposed approach provides on average a 25.84% relative WER improvement on LibriSpeech sets and rare-word evaluation compared to the baseline.

Document shadow removal is an integral task in document enhancement pipelines, as it improves visibility, readability and thus the overall quality. Assuming that the majority of practical document shadow removal scenarios require real-time, accurate models that can produce high-resolution outputs in-the-wild, we propose Laplacian Pyramid with Input/Output Attention Network (LP-IOANet), a novel pipeline with a lightweight architecture and an upsampling module. Furthermore, we propose three new datasets which cover a wide range of lighting conditions, images, shadow shapes and viewpoints. Our results show that we outperform the state-of-the-art by a 35% relative improvement in mean average error (MAE), while running real-time in four times the resolution (of the state-of-the-art method) on a mobile device.

Space-time memory (STM) network methods have been dominant in semi-supervised video object segmentation (SVOS) due to their remarkable performance. In this work, we identify three key aspects where we can improve such methods; i) supervisory signal, ii) pretraining and iii) spatial awareness. We then propose TrickVOS; a generic, method-agnostic bag of tricks addressing each aspect with i) a structure-aware hybrid loss, ii) a simple decoder pretraining regime and iii) a cheap tracker that imposes spatial constraints in model predictions. Finally, we propose a lightweight network and show that when trained with TrickVOS, it achieves competitive results to state-of-the-art methods on DAVIS and YouTube benchmarks, while being one of the first STM-based SVOS methods that can run in real-time on a mobile device.

The majority of the state-of-the-art shadow removal models (SRMs) reconstruct whole input images, where their capacity is needlessly spent on reconstructing non-shadow regions. SRMs that predict residuals remedy this up to a degree, but fall short of providing an accurate and flexible solution. In this paper, we rethink residual predictions and propose Learnable Residual Attention (LRA) and Learnable Dense Reconstruction Attention (LDRA) modules, which operate over the input and the output of SRMs. These modules guide an SRM to concentrate on shadow region reconstruction, and limit reconstruction of non-shadow regions. The modules improve shadow removal (up to 20%) and detection accuracy across various backbones, and even improve the accuracy of other removal methods (up to 10%). In addition, the modules have minimal overhead (+<1MB memory) and are implemented in a few lines of code. Furthermore, to combat the challenge of training SRMs with small datasets, we present a synthetic dataset generation pipeline. Using our pipeline, we create a dataset called PITSA, which has 10 times more unique shadow-free images than the largest benchmark dataset. Pre-training models on the PITSA significantly improves shadow removal (+2 MAE on shadow regions) and detection accuracy of multiple methods. Our results show that LRA&LDRA, when plugged into a lightweight architecture pre-trained on the PITSA, outperform state-of-the-art shadow removal (+0.7 all-region MAE) and detection (+0.1 BER) methods on the benchmark ISTD and SRD datasets, despite running faster (+5%) and consuming less memory (×150).

5G architectures will utilize the virtualization of the network functions (VNF) and the use of Multi-access edge computing (MEC) to gain multiple benefits such as simpler service orchestration, while simultaneously covering diverse use cases even with strict performance requirements. 5G service orchestration mechanisms will need to allow more efficient and flexible network deployment and operations in a resource-efficient and delay-sensitive manner. A field that is expected to be greatly boosted by these advances, is Cellular Vehicle to Everything communications. 5G will enable cooperative, connected and automated mobility services, which are often are safety critical while also having stringent delay requirements. This paper, proposes a mechanism that predicts the future position of a vehicle moving in both urban and/or highway environments. Based on this knowledge, it decides on the optimal position of VNFs so that the allocation of network resources can be preemptively requested. The objective of this mechanism is to ensure the uninterrupted, continuous connections of the vehicles, resulting in minimal or no service interruption time while ensuring an optimal utilization of Edge Cloud and MEC resources.

From smart homes to smart cities and Industry 4.0 to Transportation Systems, Internet of Things (IoT) is a domain which promises incredible growth coupled with great impact, in numerous fields. IoT networks are composed of numerous different Things, arranged in diverse topologies with diverse needs. This diversity is partially due to the numerous areas where IoT applications are utilized, which at their entirety can be referred as the IoT ecosystem. The IoT ecosystem suffers from numerous vulnerabilities, due to reasons such as design flows, hardware limitation or simply human error and is subject to various attacks targeting IoT services, platforms and networks. These attacks can have significant consequences such as economic losses, service disruption or data leaks. Cyber-attacks are an unavoidable and must be faced in tandem with the global growth of IoT networks. An approach that can assist in developing robust, intelligent Cyber-security tools for IoT is using Artificial Intelligence. In the following paper, a mechanism is presented that automatically selects appropriate mitigation actions in an optimal way to countermeasure attacks faced by IoT networks. This is achieved by using an novel Artificial Intelligence mechanism based on a Deep Neural Architecture called Pointer Networks to optimize security-related KPIs. Experimental results, show that the proposed method produces equal or better Pareto optimal solutions, performs faster compared to state-of-the-art (SoA) algorithms and scales better.

This work concerns a gait matching problem. An experimental simulation of people walking in an interior area, some of which had wearables with accelerometer and gyroscope, and some of which were recorded by some camera set up in this area, was conducted. From all devices data from single-point vectors were acquired, thus preserving privacy. The aim was to match each person carrying a wearable with the correct camera recording, if this recording existed, or not to match the person with any camera recording, if it was not recorded by a camera for sufficient time. For this matching, deliberately chosen pairs of wearable and camera features were correlated after state-of-the-art-based automatic synchronization, and the correlation matrices resulting from the feature pairs, after rationally transformed to improve results, were fused with the best linear combination. From the experiments, perfect matching results were obtained when the time series to be compared lasted at least 50sec, no matter any other assumption.

Intrusion Detection Systems (IDSs) are considered as one of the fundamental elements in the network security of an organisation since they form the first line of defence against cyber threats, and they are responsible to detect effectively a potential intrusion in the network. Many IDS implementations use flow-based network traffic analysis to detect potential threats. Network security research is an ever-evolving field and IDSs in particular have been the focus of recent years with many innovative methods proposed and developed. In this paper, we propose a deep learning model, more specifically a neural network consisting of multiple stacked Fully-Connected layers, in order to implement a flow-based anomaly detection IDS for multi-class classification. We used the updated CICIDS2017 dataset for training and evaluation purposes. The experimental outcome using MLP for intrusion detection system, showed that the proposed model can achieve promising results on multi-class classification with respect to accuracy, recall (detection rate), and false positive rate (false alarm rate) on this specific dataset.

Material aging has a significant effect on the realistic rendering of artwork objects. Small deformations of the surface structure, color or texture variations contribute to the realistic look of artwork objects. These aging effects depend on material composition, object usage, weathering conditions, and a large number of other physical, biological, and chemical parameters. In this work we focus on local deformations due to corrosion/erosion and finally cracks mainly by modeling the behavior of displacements locally. Micro-profilometry provides the quantitative measurements of the surface texture and roughness at micro-metric level, which is used to obtain information about material changes over time in terms of its surface deformation. We present a method for deriving a model for simulating aging based on micro-profilometry measurements taken on material sample plates during an emulated aging process. Subsequently, we use this model for realistic rendering of aged artwork objects.

With the proliferation of high resolution 3D scanners, the quality of recorded 3D models has greatly improved. Nonetheless, while geometric fidelity is important, color information is still required to achieve photo-realistic 3D models. In this regard, texture reconstruction techniques combine color images from several views in order to optimally color the mesh of a 3D model. Nonetheless, a major challenge that is often overlooked by existing approaches is the technical limitations of color acquisition devices that lead to erroneously colored 3D models. In this paper, a novel technique is presented that formulates texture reconstruction as an optimization problem incorporating a color correction term in its objective function. The underlying rationale is to exploit external to the 3D scanner color measurements that can be available from more reliable sensors such as a UV- VIS spectrometer. Such measurements are often available for objects of high aesthetic value such as artworks of cultural heritage objects. Through experimental evaluation of our method on a real painting we demonstrate the superiority of the proposed technique, compared to state-of-the-art texture reconstruction, providing a reliable representation of the artworks appearance both in terms of numerical accuracy and visual observation.

Attacks on the content and quality of service of IoT platforms have economic and physical consequences well beyond the Internet's lack of security. This paper describes a new research project on “Secure and Safe Internet of Things” (SerIoT) to improve both the information and physical security of IoT applications platforms in a holistic and cross-layered manner. The purpose is to be able to create secure operational IoT platfnrms for diverse applieations.

The reduction of road congestion requires intuitive urban congestion-control platforms that can facilitate transport stakeholders in decision making. Interactive ITS visual analytics tools can be of significant assistance, through their real-time interactive visualizations, supported by advanced data analysis algorithms. In this paper, an interactive visual analytics platform is introduced that allows the exploration of historical data and the prediction of future traffic through a unified interactive interface. The platform is backed by several data analysis techniques, such as road behavioral visualization and clustering, anomaly detection, and traffic prediction, allowing the exploration of behavioral similarities between roads, the visual detection of unusual events, the testing of hypotheses, and the prediction of traffic flow after hypothetical incidents imposed by the human operator. The accuracy of the prediction algorithms is verified through benchmark comparisons, while the applicability of the proposed toolkit in facilitating decision making is demonstrated in a variety of use case scenarios, using real traffic and incident data sets.

Complex multimedia data handling is as a hugechallenge for the computer science community. Social networks, manufacturing, retail, national and cyber-security, medicine, computational biology, etc. are related to a variety of heterogeneousimages and videos that create unique methodologicalchallenges. Data complexity arising from these large-scale applicationsseek for high-performance processing to obtain datainsights. Dimensionality reduction techniques reveal a managementexcellence towards multimedia data manipulation. Manydimensionality reduction techniques can be described within theframework of graph embedding, where a neighborhood graph isconstructed from the data, in order to reveal their structure. Existingtechniques usually handle only a single data representation, so-called modality, resulting in a single neighborhood graph. However, data often are multimodal, i.e. the same semantic conceptis described by various diverse representations. In the contextof graph embedding, the multiple modalities can be representedas multiple neighborhood graphs among the data. In this paper, an extension of the graph embedding framework is presented, where a multimodal graph is constructed as a weighted sumof the multiple unimodal graphs. Different from other methods, the weights of this sum are adaptively calculated as the solutionto an optimization problem, where an introduced measure ofgraph consistency, based on a rational assumption regardingthe similarities between objects, is optimized. The experimentalresults of the comparison of the proposed Multimodal GraphEmbedding (MGE) dimensionality reduction method to existingwork prove that this adaptive weighting scheme leads to superiorperformance in a number of experimental settings and datasets.

Traditional multimedia search engines retrieve results based mostly on the query submitted by the user, or using a log of previous searches to provide personalized results, while not considering the accessibility of the results for users with vision or other types of impairments. In this paper, a novel approach is presented which incorporates the accessibility of images for users with various vision impairments, such as color blindness, cataract and glaucoma, in order to rerank the results of an image search engine. The accessibility of individual images is measured through the use of vision simulation filters. Multi-objective optimization techniques utilizing the image accessibility scores are used to handle users with multiple vision impairments, while the impairment profile of a specific user is used to select one from the Pareto-optimal solutions. The proposed approach has been tested with two image datasets, using both simulated and real impaired users, and the results verify its applicability. Although the proposed method has been used for vision accessibility-based reranking, it can also be extended for other types of personalization context.

Border gateway protocol (BGP) is the main protocol used on the Internet today, for the exchange of routing information between different networks. The lack of authentication mechanisms in BGP, render it vulnerable to prefix hijacking attacks, which raise serious security concerns regarding both service availability and data privacy. To address these issues, this study presents BGPGraph, a scheme for detecting and visualising Internet routing anomalies. In particular, BGPGraph introduces a novel BGP anomaly metric that quantifies the degree of anomaly on the BGP activity, and enables the analyst to obtain an overview of the BGP status. The analyst, is afterwards able to focus on significant time windows for further analysis, by using a hierarchical graph visualisation scheme. Furthermore, BGPGraph uses a novel method for the quantification of information visualisation that allows for the evaluation, and optimal selection of parameters, in case of the corresponding visual analytics algorithms. As a result, by utilising the proposed approach, four new BGP anomalies were able to be identified. Experimental demonstration in known BGP events, illustrates the significant analytics potential of the proposed approach in terms of identifying prefix hijacks and performing root cause analysis.

Mobile devices are evolving and becoming increasingly popular over the last few years. This growth, however, has exposed mobile devices to a large number of security threats. Malware installed in smartphones can be used for a variety of malicious purposes, including stealing personal data, sending spam SMSs, and launching Denial of Service (DoS) attacks against core network components. Authentication and access-control-based techniques, employed by network operators fail to provide integral protection against malware threats. In order to solve this issue, the activity of each mobile device in the network must be taken into account, and combined with the activities of all the other devices. The communication activity in the mobile network has a source, a destination, and possibly communication weights (e.g., the number of calls between two mobile devices). This relational nature of the communication activity is naturally represented with graphs. This indicates that graphs can be utilized in order to provide better representations of the entire network activity, and lead to better detection results when compared to methods that consider the activity of each mobile device individually. Towards this end, this paper proposes a novel graph-based descriptor for the detection of anomalies in mobile networks, using billing-related information. The graph-based descriptor represents the total activity in the network. Smaller graphs are afterwards extracted from the graph-based descriptor, each one representing the activity of one mobile device (e.g., Calls or SMSs), while multiple features are calculated for each such graph. These features are subsequently used for the supervised classification on network events, and the identification of anomalous mobile devices. Experimental results and comparison of the proposed anomaly detection method to the existing work, show that the graph-based descriptor has superior performance in a variety of scenarios.

Using data visualization techniques can be of significant assistance in exploring multimedia databases. Data visualization is typically addressed as a unimodal learning task, where data are described with only one feature set, or modality. However, using multiple data modalities has been proved to increase the performance of learning methods. In this paper a novel approach for exploiting the multiple available modalities for visualization is proposed, motivated by the field of multi-objective optimization. Initially, each modality is considered separately. A graph of the dissimilarities among the data and the corresponding minimum spanning tree are formed. The suitability of a particular data placement is quantified using multiple cost functions, one for each modality. The utilized cost functions are defined in terms of graph aesthetic measures, computed for the unimodal minimum spanning trees. The cost functions are then used as the multiple objectives of a multi-objective optimization problem. Solving the problem results in a set of Pareto optimal placements, which represent different trade-offs among the various objectives. Experimental evaluation shows that the proposed method outperforms current multimodal visualization methods both in discovering more visualizations and in producing ones which are more aesthetically pleasing and easily perceivable.

This letter presents a novel probabilistic framework for augmenting the recognition performance of biometric systems with information from continuous soft biometric (SB) traits. In particular, by modelling the systematic error induced by the estimation of the SB traits, a modified efficient recognition probability can be extracted including information related both to the hard and SB traits. The proposed approach is applied without loss of generality in the case of gait recognition, where two state-of-the-art gait recognition systems are considered as hard biometrics and the height and stride length of the individuals are considered as SBs. Experimental validation on two known, large datasets illustrates significant advances in the recognition performance with respect to both identification and authentication rates.