Detail of the student project

Topic:Object recognition with incremental learning through feature adaptation
Department:Katedra kybernetiky
Supervisor:Georgios Tolias, Ph.D.
Announce as:Diplomová práce, Bakalářská práce, Semestrální projekt
Description:Traditional object classification methods train the network on images from all the object categories. In the continual learning setting, the object categories are observed sequentially i.e. the network only observes images from a single category for a given time. This leads to what is known as catastrophic forgetting where the network performance on previously seen categories drastically drops while learning to classify images from current object category. In this project we will focus on two aspects: i) storing a small amount of data (eg. Images, network features, feature statistics from different layers) from already observed object classes in a small memory buffer, and ii) how to use this buffer information to efficiently retain knowledge of previously seen object categories.
Bibliography:Iscen Zhang Lazebnik Schmid, ECCV 2020, Memory-Efficient incremental learning through feature adaptation
Responsible person: Petr Pošík