This seminar is dedicated to recent works on graph databases (models, languages, optimization, storage, standards, …), with connections with advancement in AI (graph machine learning, GNN, vector databases, …).
The seminar is informal (beside classical presentation, sessions can be dedicated to open-problems, reading sessions, techical issues, …) and on-line. Do not hesitate to contact us for a slot.
Next seminar
- Pierre Lefebvre (PhD student, ESILV)
“NeoSGG: A Scene Graph Generation Framework for Video-Surveillance Tasks”
July 2, 2024, 1pm-2pm (Oleron room A008, IRISA, and online)
Short Biography
Pierre Lefebvre is a PhD student in computer science at the ESILV engineering school in Paris la Défense. Pierre is a member of the DVRC laboratory’s “Data science, digital transformation, risks and complex systems” research group. Previously, he completed a double degree with the ENSEA engineering school and the Université du Québec à Chicoutimi (UQAC), where he obtained a master’s degree in computer science. His research topics are computer vision and graph databases. His work focuses on distributed video surveillance systems for semantic video analysis, scene understanding and complex event detection.
Summary
Video surveillance has developed considerably in recent years. Analyzing the data generated by such systems has become a major challenge. To address this issue, we propose NeoSGG, a framework for the detection of complex events in video surveillance streams using Scene Graph Generation (SGG) and pattern search queries. It is based on 1) a Deep Learning pipeline architecture for video data extraction, 2) a graph database module to efficiently structure and store detections, and 3) a querying module to interact with the generated Labeled Property Graph (LPG), enhancing the automatic analysis of scenes.
During this presentation, a brief introduction will be given to scene analysis and the notion of “Semantic Gap”. This will be followed by a general presentation of the framework and its 3-layer architecture. Then, a more in-depth presentation of the various modules and the graph data model will be provided. Examples of graphs and queries will be given to illustrate different use cases and underline the expressivity of our approach. Finally, we’ll look ahead to potential future developments. If time permits, the framework will be shown on proprietary videos to demonstrate the process of graph generation and query execution.
