In object detection, classical rule for accepting a match between a ground truth area and a detection is a 0.5 jacard ratio. But does user care about this ? Especially, can this rule be well handled by deep network training ? And if no, may user accept some relaxation of the problem if it can help the training ? In this paper, we benchmark several strategies to perform object detection with end to end deep network when metric is relaxed. Our preliminary results on different public datasets show that under this relaxations, some strategies are very facilitative to train the network outperforming the same network learned with classical strategies.