Publications
Lead: Latent realignment for human motion diffusion
Nefeli Andreou, Xi Wang, Victoria Fernández Abrevaya, Marie‐Paule Cani, Yiorgos Chrysanthou, Vicky Kalogeiton
Computer Graphics Forum, e70093, Apr 2025
This work proposes LEAD, a method that combines latent diffusion with a realignment mechanism to create a semantically structured space for generating realistic human motion from natural language. It enables both high-quality motion synthesis and motion textual inversion, outperforming modern methods in realism, diversity, and alignment with textual input.
DeepSafe: Two-level deep learning approach for disaster victims detection
Amir Azizi, Panayiotis Charalambous, Yiorgos Chrysanthou
Virtual Reality & Intelligent Hardware, 7(2), Pages 139-154, Apr 2025
We present DeepSafe, a two-level deep learning framework for disaster victim detection that combines YOLOv8 for classification and Detectron2 for precise localization. The approach improves accuracy and speed in identifying victims under challenging conditions, offering an effective tool for real-time search and rescue operations.
Demonstrating the effectiveness of combining heuristic and data-driven methods to achieve scalable and adaptive motion styles
Amir Azizi, Panayiotis Charalambous, Yiorgos Chrysanthou
Presented at: 2025 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), Mar 2025
We propose a hybrid motion style transfer framework that combines heuristic methods (e.g., RBF and phase features) with data-driven autoencoders to balance efficiency and adaptability. The approach reduces dataset requirements while maintaining high-quality results, demonstrating scalable and flexible motion generation across multiple datasets.
A novel multidisciplinary approach for reptile movement and behavior analysis
Savvas Zotos, Marilena Stamatiou, Sofia-Zacharenia Marketaki, Duncan J. Irschick, Jeremy A. Bot, Andreas Aristidou, Emily L. C. Shepard, Mark D. Holton, Ioannis N. Vogiatzakis
Integrative Zoology, Feb 2025
This paper introduces a multidisciplinary approach to studying reptile behavior, combining tri-axial accelerometers, video recordings, motion capture systems, and 3D reconstruction to create detailed digital archives of movements and behaviors. Using two Mediterranean reptiles as case studies, it highlights the potential of this method to advance research on complex and understudied behaviors, offering ecological insights and tools for behavioral analysis.
Deep convolutional generative adversarial networks in retinitis pigmentosa disease images augmentation and detection
Paweł Powroźnik, Maria Skublewska-Paszkowska, Katarzyna Nowomiejska, Andreas Aristidou, Andreas Panayides, Robert Rejdak
Advances in Science and Technology Research Journal, Volume 19, no. 2, pages 321-340., Nov 2024
This study leverages Deep Convolutional Generative Adversarial Networks (DCGAN) and hybrid VGG16-XGBoost techniques to enhance medical datasets, focusing on retinitis pigmentosa, a rare eye condition. The proposed method improves image clarity, dataset augmentation, and detection accuracy, achieving over 90% in key performance metrics and a 19% increase in baseline classification accuracy.
Identifying and Animating Movement of Zeibekiko Sequences by Spatial Temporal Graph Convolutional Network with Multi Attention Modules
Maria Skublewska-Paszkowska, Paweł Powroźnik, Marcin Barszcz, Krzysztof Dziedzic, Andreas Aristidou
Advances in Science and Technology Research Journal, Volume 18, no. 8, pages 217-227., Nov 2024
This study employs optical motion capture technology to document and translate the Zeibekiko dance into a 3D virtual environment. Using a Spatial Temporal Graph Convolutional Network with Multi Attention Modules (ST-GCN-MAM), the system accurately captures and classifies essential dance sequences by focusing on key body regions, enabling precise, realistic virtual animations with applications in gaming, video production, and digital heritage preservation.
Underwater Virtual Exploration of the Ancient Port of Amathus
Andreas Alexandrou, Filip Skola, Dimitrios Skarlatos, Stella Demesticha, Fotis Liarokapis, Andreas Aristidou
Journal of Cultural Heritage, Volume 70, pages 181-193, November–December 2024., Sep 2024
This work focuses on the digital reconstruction and visualization of underwater cultural heritage, providing a gamified virtual reality (VR) experience of Cyprus' ancient Amathus harbor. Utilizing photogrammetry, our immersive VR environment enables seamless exploration and interaction with this historic site. Advanced features such as guided tours, procedural generation, and machine learning enhance realism and user engagement. User studies validate the quality of our VR experiences, highlighting minimal discomfort and demonstrating promising potential for advancing underwater exploration and conservation efforts.
Design and Implementation of an Interactive Virtual Library based on its Physical Counterpart
Christina-Georgia Serghides, Giorgos Christoforidis, Nikolas Iakovides, Andreas Aristidou
Virtual Reality, Volume 28, Article Number 124, June, 2024, Springer., Jun 2024
This work explores the creation of a digital replica of a physical Library, using photogrammetry and 3D modelling. A Virtual Reality (VR) platform was developed to immerse users in a virtual library experience, which can also serve as a community and knowledge hub. A perceptual study was conducted to understand the current usage of physical libraries, examine the users’ experience in VR, and identify the requirements and expectations in the development of a virtual library counterpart. Five key usage scenarios were implemented, as a proof-of-concept, with emphasis on 24/7 access, functionality, and interactivity. A user evaluation study endorsed all its key attributes and future viability.
Improving Image Reconstruction using Incremental PCA-Embedded Convolutional Variational Auto-Encoder
Amir Azizi, Panayiotis Charalambous, Yiorgos Chrysanthou
Presented at: The 32nd International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision (WSCG 2024), May 2024
We propose an efficient image reconstruction method using Convolutional Variational Autoencoders enhanced with Incremental PCA to better capture key latent features. The approach improves image quality and scalability while reducing processing time, demonstrating strong performance on MNIST.
Behavioral Landmarks: Inferring Interactions from Data
Marilena Lemonari, Panayiotis Charalambous, Andreas Panayiotou, Yiorgos Chrysanthou, Julien Pettré
Presented at: Eurographics 2024 Posters, May 2024
Weaim to unravel complex agent-environment interactions from trajectories, by explaining agent paths as combinations of predefined basic behaviors. We detect trajectory points signifying environment-driven behavior changes, ultimately disentangling interactions in space and time; our framework can be used for environment synthesis and authoring, shown by our case studies.
Overcoming Challenges of Cycling Motion Capturing and Building a Comprehensive Dataset
Panayiotis Kyriakou, Marios Kyriakou, Yiorgos Chrysanthou
Presented at: The Creating Lively Interactive Populated Environments (CLIPE 2024) Workshop, The Eurographics Association, Apr 2024
This article outlines a methodology for capturing cyclist motion using motion capture (mocap) hardware and creating a publicly available comprehensive dataset. It features a modular system with innovative marker placement, and the resulting dataset is used to produce 3D visualizations and various data representations, which are shared in an online library for public access and collaborative research.
LexiCrowd: A Learning Paradigm towards Text to Behaviour Parameters for Crowds
Marilena Lemonari, Nefeli Andreou, Nuria Pelechano, Panayiotis Charalambous, Yiorgos Chrysanthou
Presented at: The Creating Lively Interactive Populated Environments (CLIPE 2024) Workshop, The Eurographics Association, Apr 2024
This work uses a pre-trained Large Language Model (LLM) to generate pseudo-pairs of text and behavior labels, then trains a variational auto-encoder (VAE) on this synthetic dataset, constraining the latent space with a latent label loss for interpretable behavior parameters. Our model, tested with human-provided textual descriptions of crowd datasets, can parameterize unseen sentences to produce novel behaviors compatible with simulator parameters, demonstrating its potential for text-to-crowd generation and full sentence generation from behavior profiles.
Virtual Instrument Performances (VIP): A Comprehensive Review
Theodoros Kyriakou, Mercè Álvarez, Andreas Panayiotou, Yiorgos Chrysanthou, Panayiotis Charalambous, Andreas Aristidou
Computer Graphics Forum, Volume 43, Issue 2, Apr 2024
Presented at: Eurographics 2024, EG'24 STAR papers
The evolving landscape of performing arts, driven by advancements in Extended Reality (XR) and the Metaverse, presents transformative opportunities for digitizing musical experiences. This comprehensive survey explores the relatively unexplored field of Virtual Instrument Performances (VIP), addressing challenges related to motion capture precision, multi-modal interactions, and the integration of sensory modalities, with a focus on fostering inclusivity, creativity, and live performances in diverse settings.
Digitizing Traditional Dances Under Extreme Clothing: The Case Study of Eyo
Temi Ami-Williams, Christina-Georgia Serghides, Andreas Aristidou
Journal of Cultural Heritage, Volume 67, pages 145–157, February 2024., Mar 2024
Presented at: The International Council for Traditional Music (ICTM) 2023 and the Cyprus Dance Film Festival (CDFF) 2023
This work examines the challenges of capturing movements in traditional African masquerade garments, specifically the Eyo masquerade dance from Lagos, Nigeria. By employing a combination of motion capture technologies, the study addresses the limitations posed by "extreme clothing" and offers valuable insights into preserving cultural heritage dances. The findings lead to an efficient pipeline for digitizing and visualizing folk dances with intricate costumes, culminating in a visually captivating animation showcasing an Eyo masquerade dance performance.
SparsePoser: Real-time Full-body Motion Reconstruction from Sparse Data
Jose Luis Pontón, Haoran Yun, Andreas Aristidou, Carlos Andújar, Nuria Pelechano
ACM Transactions on Graphics, Volume 43, Issue 1, Article No.: 5, pages 1–14., Oct 2023
Presented at: SIGGRAPH Asia 2023.
SparsePoser is a novel deep learning-based approach that reconstructs full-body poses using only six tracking devices. The system uses a convolutional autoencoder to generate high-quality human poses learned from motion capture data and a lightweight feed-forward neural network IK component to adjust hands and feet based on the corresponding trackers.