Ioannis Pitas’s research while affiliated with Aristotle University of Thessaloniki and other places

... Advancements in machine intelligence and autonomous systems have dramatically fueled the integration of environmental perception technologies into daily life and various industries [1][2][3][4][5]. This widespread adoption is prominently seen in applications such as autonomous cars [6], smart wheelchairs [7], and unmanned ground vehicles [8]. Recently, researchers have shifted their focus toward enhancing both driving safety and comfort [9,10]. ...

Reference:

A glance over the past decade: road scene parsing towards safe and comfortable autonomous driving

... Deep learning has made remarkable advancements in fields such as autonomous driving [1]- [3], with real-time object detection technologies, exemplified by YOLOs and DETRs, gaining widespread adoption [4], [5]. However, due to factors such as light attenuation, color distortion, and difficulty distinguishing targets from coral reefs, mud, and other underwater structures, the development of highperformance real-time underwater object detection (UOD) has been relatively slow [6]. ...

Reference:

You Sense Only Once Beneath: Ultra-Light Real-Time Underwater Object Detection

... It is widely accepted that there are three major components impacting inference performance: Efficiency, Consistency and Accuracy [42][43][44]. Efficiency refers to how quickly the model processes and returns responses. It is influenced by (a) Response Time (i.e., the time taken for the model to generate a response after receiving a query); (b) Server Busy Rate (SBR, i.e., the proportion of queries that cannot be processed due to system overload or resource constraints). ...

Reference:

Evaluating Large Language Models in Code Generation: INFINITE Methodology for Defining the Inference Index

... Neural networks are a powerful and flexible tool for forecasting [19][20][21]. When determining what exactly should be predicted, it is necessary to specify the variables that are analyzed and used in the forecasting process. ...

Reference:

MODELS FOR PREDICTING CHANGES IN PUBLIC OPINION DURING THE IMPLEMENTATION OF THE NARRATIVE IN SOCIAL MEDIA

... Patsiouras et al. [3] ...

Reference:

EXPLORING WORD EMBEDDINGS FOR SENTIMENT ANALYSIS OF MARATHI POLITICAL TWEETS: A MACHINE LEARNING APPROACH

... A few leading countries have been incorporating UAV swarms in their military attacks [6,7] and they have been proven to be extremely effective. UAV swarms have also been used in the film industry for cinematography purposes [8,9] and in the entertainment sector to create appealing drone shows that grab the attention of tourists [10,11], such as in Dubai. The practical implications of securing UAV operations extend across diverse sectors, showcasing the transformative potential of this technology when effectively safeguarded against threats such as GPS spoofing. ...

Reference:

Enhancing UAV Security Against GPS Spoofing Attacks Through a Genetic Algorithm-Driven Deep Learning Framework

... A typical stage where the NMS are struggle in performing are, when they are operated on the image, depicting the objects in a complex scenes, and in more of complex and several occlusions [17]. To overcome all these complexities faced by the existing approaches, such as the complexity in background, increased computational time, less accuracy rates, intensive calculation protocols [18][19][20], the proposed model used the Deep Pliable YOLOv5 model, for the aircraft detection using the aircraft and airbus dataset, using the Adaptive Spatial Pooling layer for obtaining the features via a parallelised method, which are more advanced than the traditional pooling method in the image recognition. ...

Reference:

Detection and recognition of aircraft vehicle-A supple approach using deep pliable YOLOv5

... Parallel implementations have significantly improved computational efficiency on embedded GPU architectures, demonstrating more than 40 times speedups while maintaining detection accuracy [29]. Neural attention-driven techniques have emerged to better handle occlusions in pedestrian detection by jointly processing geometric and visual properties through sequenceto-sequence formulations [30]. For overlapping object scenarios, adaptive NMS strategies dynamically adjust IoU thresholds based on object density, substantially enhancing performance for weed detection applications [31]. ...

Reference:

Differentiable NMS via Sinkhorn Matching for End-to-End Fabric Defect Detection

... Zhang et al. proposed a multidimensional scaling method based on the Wasserstein-Fourier distance to classify complex time series from a frequency domain perspective . A fast multidimensional scaling method on big geospatial data using neural networks is presented by Mademlis et al., where sampling a small subset of the original dataset is conducted (Mademlis et al., 2023). Multidimensional scaling is also combined with principal component analysis to analyze varietal sedimentary provenance data (Vermeesch et al.,2023), and it is used to investigate and compare the similarity of writing prompts in the IELTS and TOEFL iBT tests (Khademi, 2023). ...

Reference:

A novel clustering method for complex signals and feature extraction based on advanced information-based dissimilarity measure

... In early approaches, the summarization component was composed of LSTM units that estimated the frames' importance according to their temporal dependence (thus indicating the most significant video parts for inclusion in the summary), while the reconstruction of the video based on the specified summary was performed using trainable auto-encoders (Mahasseni et al., 2017;Apostolidis et al., 2019;Yuan et al., 2020) that in some cases were combined with tailored attention mechanisms (Jung et al., 2019;Apostolidis et al., 2020;Kanafani et al., 2021). In more recent methods, the selection of the most important frames or fragments for the summary was assisted by trainable Actor-Critic models (Apostolidis et al., 2021a;Alexoudi et al., 2023), self-attention mechanisms (He et al., 2019;Jung et al., 2020;Liang et al., 2022), spatio-temporal networks (Wu et al., 2021) or knowledge distillation mechanisms (Sreeja & Kovoor, 2022). A less popular approach for unsupervised video summarization is based on the definition of hand-crafted reward functions about specific properties of the generated summary, and the use of the computed rewards for training video summarization architectures based on reinforcement learning. ...

Reference:

Explainable Video Summarization for Advancing Media Content Production