Me

Alexandros Stergiou

Utrecht University Computer Science Ph.D. student
a.g.stergiou@uu.nl

Bio

My work is focused on Computer Vision and Machine Learning

I am currently a full-time PhD student at Utrecht University's Department of Information and Computing Sciences. My project is based on Machine Learning and Computer Vision for human action and interaction recognition in everyday social settings, supervised by Dr. Ronald Poppe and Prof. Remco Veltkamp. I am also working on improvements on the efficiency and explainabilty of spatio-temporal Deep Learning and Artificial Intelligence methods, in scene understanding and classification of videos.

Before my PhD, I obtained my Masters in Advance Computer Science from University of Essex working with Dr. Adrian Clark. During this time I also worked at the Institute for Analytics and Data Science (IADS) for a summer project on Neural Networks for structured population data with Dr. Spyros Samothrakis. My Bachelor was also at the University of Essex in Computer Science.

PUBLICATIONS

PUBLICATIONS LIST

A full list can be found in my GoogleScholar profile.


learn2cycle srtg-active
Learn to cycle: Time-consistent feature discovery for action recognition

Alexandros Stergiou and Ronald Poppe

Pattern Recognition Letters, 2021

class-reg human-human-active
Learning Class-Specific Features with Class Regularization for Videos

Alexandros Stergiou, Ronald Poppe, Remco Veltkamp

Applied Sciences, 2020

feature-pyramids human-human-active
Class Feature Pyramids for Video Explanation

Alexandros Stergiou, Georgios Kapidis, Grigorios Kalliatakis, Christos Chrysoulas,
Ronald Poppe, Remco Veltkamp

Intenrantional Conference on Computer Vision Workshop (ICCVW), 2019

FAST_conv
Spatio-Temporal FAST 3D Convolutions for Human Action Recognition

Alexandros Stergiou, Ronald Poppe

Intenrantional Conference on Machine Learning Applications (ICMLA), 2019

saliency-tubes human-human-active
Saliency Tubes: Visual Explanations for Spatio-Temporal Convolutions

Alexandros Stergiou, Georgios Kapidis, Grigorios Kalliatakis, Christos Chrysoulas,
Remco Veltkamp, Ronald Poppe

International Confernce on Image Procesing (ICIP), 2019

h2h_survey
Analyzing human-to-human interactions: a survey

Alexandros Stergiou, Ronald Poppe

Computer Vision and Image Understanding, 2019

conceiving_human
Conceiving Human Interaction by Visualising Depth Data of Head Pose Changes and Emotion Recognition via Facial Expressions

Grigorios Kalliatakis, Alexandros Stergiou, Nikolaos Vidakis

Computers, 2017

TEACHING

MODULES THAT I HAVE BEEN INVOLVED IN

Fig1
INFOMCV (2019-2020) - Computer Vision (Master Course)

Lecturer for the Deep Learning lectures

Fig1
INFOMCV (2018-2019) - Computer Vision (Master Course)

Lecturer and Instructor for the Deep Learning/Neural Networks lectures

Fig1
INFOMCV (2017-2018) - Computer Vision (Master Course)

Instructor for the Deep Learning/Neural Networks lectures

PROJECTS

PROJECTS WORKBENCH

These are some of my latest projects and code.


Video dataset SQL converter

A package made as a solution to the problem of using video inputs in Machine Learning models. As extracting and storing frames in .JPEG/.PNG files will quickly increase the memory requirements and more importantly the number of inodes. This package provides a convenient alternative. Video frames are stored as blobs at database files .db which can be read as quickly as the .JPEG files but without the additional large memory requirements.
 code

Depth-wise 3D Convolutions for Keras

An extension of separable convolutions for 3D volumes. Performs volumetric convolutions for each channel of the input volume and will increase the output volume based on the number of convolutional operations.
 code

Bootstrapping for polling data from the 2016 EU referendum vote in the United Kingdom

This project presents the use of Bootstrapping (or also known as bagging), to normalise the features of the polling population in order to represent the entire population that is eligible to vote. Additionally, for representing connections between individual features during bootstrapping, a nearest neighbour method is employed in order for the newly created examples to resemble others from the general population. This technique will be applied to data from the 2016 Referendum that held the question of the remain or exit of the United Kingdom from the European Union.
 PDF code

PERSONAL INFO

PRESENT & PAST AFFILIATIONS




© Alexandros Stergiou