Συντάχθηκε 10-01-2024 12:31
Τόπος: Λ - Κτίριο Επιστημών/ΗΜΜΥ, 145Π-58
Έναρξη: 17/01/2024 11:00
Λήξη: 17/01/2024 12:15
Abstract
Retrieving videos from the same incident can be formulated as a search-by-example problem aiming to discover all videos in a database related to a given query. Hence, to tackle this problem, the following has to be specified: (i) what videos are considered related? (ii) how do we measure similarity between two videos to determine relevance? In this talk, we will delve into our advancements in the video retrieval field to provide answers to both questions. First, we will go through our definitions regarding the same incident videos and the composition of a large-scale dataset of user-generated videos that simulate the problem and cover its benchmarking needs. Then, we will review our proposed approaches for the estimation of video similarity. They can be roughly classified into three categories: (i) Coarse-grained approaches that extract global video representations combined with simple similarity metrics. Our solutions leverage deep Convolutional Neural Networks and Deep Metric Learning to extract global video representations and map videos to feature spaces that preserve video relations. (ii) Fine-grained approaches that employ spatio-temporal representations and similarity functions. Our approach involves a video similarity learning network that captures fine-grained relations between videos, trained with supervised and self-supervised methodologies. (iii) Re-ranking approaches that combine methods from the two previous categories to perform more efficient search. We devise a knowledge distillation scheme that trains two student networks based on a teacher, which are combined with a selector network tuned to achieve an optimal performance-efficiency trade-off. Finally, we will benchmark the proposed methods on the composed dataset and compare them with several state-of-the-art approaches.
About the Speaker
Dr. Giorgos Kordopatis-Zilos is a postdoctoral researcher with the Czech Technical University in Prague and part of the Visual Recognition Group (VRG). Prior to that, he was a postdoctoral researcher in the Information Technologies Institute (ITI) of the Centre for Research and Technology Hellas (CERTH) with the Media Analysis, Verification and Retrieval (MeVer) group. In 2013, he received his diploma in electrical and computer engineering from the Aristotle University of Thessaloniki (AUTH), Greece. In 2021, he was awarded his Ph.D. from the Queen Mary University of London, UK. His research interests include similarity learning, multimedia retrieval and matching, supervised and self-supervised learning, knowledge distillation, location estimation, deepfake detection, and forgery localization.
Σύνδεσμος εκδήλωσης
