Public Thesis defense - ICTEAM

Stream Based Active Distillation and Holonic Architecture for Scalable Camera Systems by Dani Manjah

This thesis addresses the technical and financial challenges of deploying lightweight neural networks for large-scale video analytics. We posit that these models should be specialized for their local conditions which can be accomplished by collecting and training on data from individual cameras. However, this approach requires recurrent annotation of images for each camera and potentially a unique model per camera.

We propose first to employ a large, general-purpose model to provide automatic pseudo-annotations, thereby reducing the reliance on costly human-based labelling. By selecting only, the images for which the small model shows high confidence, this framework reduces the amount of data needed for training, which in turn lowers both computational and pseudo-annotation costs.

Next, to reduce the number of models needed in a multi-camera setup, we group cameras by similarity and train one model per group. By controlling the size of these clusters, we balance the lightweight models’ ability to learn targeted distributions with the need for data diversity.

Finally, we propose an architecture that treats each camera as part of a network of semi-independent, interlinked units. This structure allows the network to adapt seamlessly as cameras are added or removed, without requiring extensive reconfiguration.

These advancements lay the groundwork for more scalable, edge-based video monitoring, maintaining accuracy with direct industrial applications.

Jury’s members:

Prof. Benoit Macq (UCLouvain), supervisor

Prof. Stéphane Galland (UTBM, France), supervisor

Prof. David Bol (UCLouvain), chairperson

Prof. Christophe De Vleeschouwer (UCLouvain), secretary

Prof. Manuel Kolp (UCLouvain)

Prof. Frédéric Dufaux (Supélec, France)

Prof. Marc Van Drooghenbroeck (ULiège)

Pay attention:

The public defense of Dani Manjah scheduled for Monday November 18 at 03:30p.m. will also take place in the form of a video conference.

Télécharger l’annonce.