The Cifre thesis of Van Huy Vo was supervised by Patrick Perez from Valeo.ai and Jean Ponce, PR[AI]RIE Scientific Director and chair holder.
Here is a short abstract of the thesis: Object detectors are important components in many intelligent systems. They are typically obtained with fully-supervised training which require costly annotated data. This thesis considers annotation-efficient alternatives to fully-supervised object detection. In particular, we discuss several approaches to Unsupervised Object Discovery, which aims to link images containing similar objects and localize these without any supervision. We also introduce a framework which combines active learning and weakly-supervised object detection that demonstrates a better detection performance / annotation cost trade-off than both fully- and weakly-supervised object detection.