The document presents a method for improving object detection by acquiring localization confidence from predicted intersection-over-union (IOU) between predicted and ground-truth bounding boxes. It addresses two problems: 1) accurate boxes are eliminated in non-maximum suppression due to using classification confidence rather than IOU, and 2) bounding box regression can degrade localization with multiple applications. The proposed method uses an IOU prediction network on transformed regions of interest, a differentiable PrRoI pooling layer, and IOU-guided non-maximum suppression. Experimental results show improved precision and recall over alternative approaches.