Agnostic object counting with deep learning models

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enObject counting is a fundamental computer vision task with applications ranging from industrial quality control and traffic monitoring to agriculture and medical image analysis. While recent deep learning approaches have significantly improved counting performance, many state-of-the-art methods still rely on large annotated datasets and task-specific training procedures, limiting their scalability and generalization capabilities across object categories. This thesis investigates whether class-agnostic object counting can be performed solely from visual representations extracted by a self-supervised Vision Transformer, without relying on annotated training data or additional learning stages. More specifically, we study the extent to which the features learned by a self-supervised vision backbone can support a robust and generalizable counting framework, capable of handling instance-scale variations, appearance changes, and occlusions while remaining applicable across previously unseen object categories.