Testing ground-truth errors in an automotive dataset for a DNN-based object detector