PhD Defense: 'Object Detection with Minimal Supervision: From Few-Shot to Open- Vocabulary'

Object detection has achieved remarkable progress with deep learning, yet remains heavily reliant on large annotated datasets, fixed category vocabularies, and stable domains. This thesis explores object detection under minimal supervision, progressively relaxing these constraints across three settings.  

First, in few-shot detection, we propose a pseudo-label mining framework to expand training data with high-quality, diverse annotations from unlabeled images. Second, in cross-modal open-vocabulary detection, we introduce a training-free adaptation pipeline for transferring RGB-trained models to X-ray imagery, supported by a new benchmark. Finally, in the category-free scenario, we develop a caption-driven approach for discovering and grounding novel categories without predefined taxonomies.  

Together, these contributions advance towards adaptable, data-efficient detection systems capable of operating in diverse, open-world environments.

Supervisor: Manuel Mucientes Molina