Event
Last Fridays Talks: Fine-Grained analysis
Location
Date
Talk 1
Insights into Generation and Analysis of Faces
Abstract
Facial expressions provide valuable information about emotions and even underlying mental health conditions. Small changes in facial features can offer insights across various domains, including diagnostics and behavioral analysis.
While modern diffusion models can generate photorealistic images, precise and disentangled facial editing remains a challenge.
This talk presents selected methods for discovering semantic directions in the latent space of diffusion models, enabling fine-grained facial edits without requiring architectural modifications or additional model fine-tuning. By leveraging both supervised and unsupervised approaches, these techniques enable interpretable and disentangled manipulations of attributes such as pose, expression, and identity.
Understanding these latent directions provides a promising avenue for many meaningful applications in human-centered AI and beyond.
Speaker
Bio
Stella Graßhof is an Assistant Professor in the Data Science section at the IT University of Copenhagen (ITU), and the pioneercentre for AI. Her research centers on human-centered machine learning, with a focus on generative AI in the visual domain. She leverages images and videos for the analysis and synthesis of human faces and motion, aiming to enhance diagnostics, personalize healthcare, and improve accessibility.
Talk 2
Visual Localization in Underwater Robotics: Deep Learning and 3D Perception Challenges
Abstract
Visual localization is essential for autonomous navigation in underwater robotics, where the absence of GPS and harsh visual conditions pose significant challenges for traditional methods.
This talk explores deep learning architectures for visual localization, highlighting both the opportunities and unique challenges they present in the underwater domain.
A central focus will be the problem of high-dimensional data representation in 3D perception, and the need for neural architectures specifically designed to handle this complexity.
We will also discuss strategies to overcome the scarcity of real-world data, including the use of synthetic datasets and dimensionality-aware model design.
Speaker
Bio
Olaya Alvarez-Tuñón is a Computer Vision and Deep Learning Engineer specializing in Visual SLAM (Simultaneous Localization and Mapping) for underwater robotics.
She holds a PhD from Aarhus University through the Marie Curie ITN project REMARO, where she focused on deep-learning-based localization in challenging environments.
Currently, she works at EIVA A/S and ITU Copenhagen on the DeepODO project, where she continues her research on combining geometric reasoning with deep learning to tackle the unique challenges of underwater visual SLAM.