MPDIoU: A Loss for Efficient and Accurate Bounding Box Regression

• DOI: 10.48550/arxiv.2307.07662
• Publication Date: 2023-01-01
• Authors: Siliang Ma; Yong Xu
• Keywords: Minimum bounding box; Bounding overwatch; Pascal (unit); Rectangle

Bounding box regression (BBR) has been widely used in object detection and instance segmentation, which is an important step localization. However, most of the existing loss functions for bounding cannot be optimized when predicted same aspect ratio as groundtruth box, but width height values are exactly different. In order to tackle issues mentioned above, we fully explore geometric features horizontal rectangle propose a novel similarity comparison metric MPDIoU based on minimum point distance, contains all relevant factors considered functions, namely overlapping or non-overlapping area, central points deviation height, while simplifying calculation process. On this basis, function MPDIoU, called LMPDIoU . Experimental results show that applied state-of-the-art segmentation (e.g., YOLACT) YOLOv7) model trained PASCAL VOC, MS COCO, IIIT5k outperforms functions.