There are a lot of works aiming to reduce the need of human annotations for object detection: self supervised training, interactive verification instead of annotation or weakly supervised training. For example, only pointing object centres is a faster to annotate but weaker ground truth than providing bounding boxes or detailed segmentation mask. Although not usable for large areas such as roads, vegetation and buildings, centers can be used to learn adequate detectors and segmentors. We perform a comparative analysis on four public remote sensing datasets on the task of vehicle detection and show that centre annotations is a competitive baseline compared to other more sophisticated annotations.