Unlike classical detectors trained on a closed set of categories, and VLMs (vision-language models) that are heavy, non-explainable, and limited to a fixed vocabulary, this thesis proposes an explainable and frugal approach to open-world object detection. Indeed, object detectors in an open world are capable not only of recognizing known objects, but also of signaling unknown or unexpected objects.

The thesis project relies on using VLMs solely as offline teachers during training in order to distill their knowledge into a lightweight, real-time detector. By leveraging learning and reasoning from a hierarchy of shared visual concepts (covering appearance, attributes, and semantic families), the proposed model will be able not only to recognize known objects, but also to signal and characterize unknown objects without prior prompts. This approach will thus provide visual explanations directly grounded in the image, meeting the demands for robustness, transparency, and frugality essential in applications across numerous safety-critical domains (mobility, defense and security, healthcare, and industrial inspection).