Nikolaos Kilis

Aristotle University of Thessaloniki (A.U.Th.)

July 2021

MSc in Artificial Intelligence, with Distinction Modules (Average Grade 87.3%): Machine Learning (75%), Intelligent Systems Programming (75%), Computer Vision (100%), Computational Intelligence (90%), Advanced Machine Learning (95%), Deep Learning (88%), Automated Planning,Scheduling and Constraint Satisfaction (80%), Natural Language Processing (95%).
Master Thesis: "Human Action Recognition with Graph Neural Networks". A multi-class classification between graphs that consist of human skeleton joints, extracted from video frames using Deep Learning frameworks. Thessaloniki, Greece

Democritus University of Thrace (D.U.Th.)

July 2018

Diploma in Electrical and Computer Engineering Diploma Thesis: “Automatic dereverberation of recordings”. A novel method for removing noise, echo and reverberation in audio signals. Xanthi, Greece

IEEE International Conference on Image Processing 2025

Feb. 2025

A rapid prototyping framework for customizable low-poly 3D assets Introduction of an AI-driven pipeline designed to facilitate scalable and memory-efficient 3D content creation. Our approach integrates 3D reconstruction, mesh simplification and adjustment, texture enhancement, semantic enrichment, and customizable stylization into a cascade of AI methods. Additionally, the pipeline incorporates a human-in-the-loop approach for texture synthesis, enabling efficient parameter tuning to minimize manual effort while improving overall synthesis quality. As is experimentally shown, this methodology not only optimizes parameter-rich systems but also provides a scalable framework for automated and user-driven refinement, making complex 3D content creation more accessible and efficient.

Digital Applications in Archaeology and Cultural Heritage journal

Dec. 2024

AI tools for generating Digital Heritage Twins enhancing storytelling in educational games Introduction of a (semi-) automated user-friendly framework for 3D asset acquisition, enhancement, and semantic enrichment. These include computer vision modules developed for super-resolution and style transfer on 2D video frames, which are employed to generate 3D assets via multi-view reconstruction. The proposed framework further supports an Active Learning (AL) process that produces images from novel viewing angles inside virtual environments. These views are exploited to improve previously generated 3D assets. 3D content creators can utilize the proposed framework to digitize new or modify and enhance existing assets regardless of size or shape with minor editing from a single video.

IEEE 3rd International Conference on Intelligent Reality (ICIR)

Dec. 2024

AI services for generating customizable game assets Introduction of user-friendly AI services designed for the enhancement and semantic enrichment of 2D images, videos, and 3D assets, focusing on items of historical importance. 3D asset acquisition via 3D reconstruction from an input video is also offered as a service that may be used in conjunction with the other services. By capturing an object of interest in a short video and by specifying a few parameters to dictate a desired output, game developers can utilize these services to create customizable assets that promote user immersion.

Virtual Reality, Springer publisher

Dec. 2023

A Real-Time Wearable AR System for Egocentric Vision on the Edge Introduction of the DARLENE wearable AR system, analysing its specifications, overall architecture and core algorithmic components. The system comprises of a wearable computing node responsible for real-time analysis of dynamic scenes supporting functionalities for instance segmentation, tracking and pose estimation. To improve user experience, a novel approach is proposed for the adaptive rendering of AR content by considering the user’s stress level, the context of use, and the environmental conditions for adjusting the level of presented information towards enhancing their situational awareness. https://www.researchgate.net/publication/364256143_A_Real-Time_Wearable_Ar_System_for_Egocentric_Vision_on_the_Edge

Springer Nature Switzerland publisher

Sep. 2023

Augmentation Based on Artificial Occlusions for Resilient Instance Segmentation An augmentation methodology that allows for sufficient and more balanced model training relying on a small number of annotated data. Additionally, two new datasets are introduced for instance segmentation including the semantic class firearm, for security applications. https://link.springer.com/chapter/10.1007/978-3-031-43153-1_4

IEEE publisher

Oct. 2022

An Efficient Framework for Human Action Recognition Based on Graph Convolutional Networks Introducing a framework for increasing Human Action Recognition (HAR) performance of Graph Convolutional Network based methods. A novel weight initialization technique together with a missing-joint-handling pre-processing step, indicated favorable scores for both 2D and 3D skeleton-based HAR datasets. https://ieeexplore.ieee.org/abstract/document/9897258

MDPI publisher

Sep. 2019

A Novel Scheme for Single-Channel Speech Dereverberation Based on the aforementioned Diploma Thesis: “Automatic dereverberation of recordings”. A paper introducing a two-stage single-channel speech enhancement scheme, tested on both simulated and real reverberant audio signals. https://www.mdpi.com/2624-599X/1/3/42