Towards Few-Annotation Learning for Object Detection: Are Transformer-based Models More Efficient ?