In object detection with deep neural networks, the box-wise objectness score\ntends to be overconfident, sometimes even indicating high confidence in\npresence of inaccurate predictions. Hence, the reliability of the prediction\nand therefore reliable uncertainties are of highest interest. In this work, we\npresent a post processing method that for any given neural network provides\npredictive uncertainty estimates and quality estimates. These estimates are\nlearned by a post processing model that receives as input a hand-crafted set of\ntransparent metrics in form of a structured dataset. Therefrom, we learn two\ntasks for predicted bounding boxes. We discriminate between true positives\n($\\mathit{IoU}\\geq0.5$) and false positives ($\\mathit{IoU} < 0.5$) which we\nterm meta classification, and we predict $\\mathit{IoU}$ values directly which\nwe term meta regression. The probabilities of the meta classification model aim\nat learning the probabilities of success and failure and therefore provide a\nmodelled predictive uncertainty estimate. On the other hand, meta regression\ngives rise to a quality estimate. In numerical experiments, we use the publicly\navailable YOLOv3 network and the Faster-RCNN network and evaluate meta\nclassification and regression performance on the Kitti, Pascal VOC and COCO\ndatasets. We demonstrate that our metrics are indeed well correlated with the\n$\\mathit{IoU}$. For meta classification we obtain classification accuracies of\nup to 98.92% and AUROCs of up to 99.93%. For meta regression we obtain an $R^2$\nvalue of up to 91.78%. These results yield significant improvements compared to\nother network's objectness score and other baseline approaches. Therefore, we\nobtain more reliable uncertainty and quality estimates which is particularly\ninteresting in the absence of ground truth.\n