We need to train our detection models with rotated samples, because for now we support rotation but our models are not fitted to predict rotated boxes, which yields to very poor results.