Annotation-efficient learning for object discovery and detection. (Les méthodes d'apprentissage efficaces en annotation pour la découverte et la détection d'objets)

Huy V. Vo

Abstract: Object detectors are important components of intelligent systems such as autonomous vehicles or robots. They are typically obtained with fully-supervised training, which requires large manually annotated datasets whose construction is time-consuming and costly. This thesis studies alternatives to fully-supervised object detection that work with less or even no manual annotation. We focus in the first part of this thesis on the unsupervised object discovery problem, which, given an image collection without manual annotation, aims at identifying pairs of images that contain similar objects and localizing these objects. We discuss two optimization-based approaches(OSD and rOSD), a ranking method (LOD) and a simple seed-growing approach that exploits features from self-supervised transformers (LOST) to this problem. In the second part of the thesis, we consider a practical scenario which combines weakly-supervised and active learning for training an object detector, and propose BiB, an active learning strategy tailored for this scenario. We show that our pipeline offers a better trade-off between annotation cost and effectiveness than both weakly- and fully-supervised object detection.