Optimization of the Bounding Box Regression Process of SSD Model

摘要

Intersection over Union (IoU) has always been the most popular evaluation metric used in object detection benchmarks. However, IoU has a disadvantage that it is not feasible to optimize without overlapping bounding boxes. Therefore, proposed a generalized version as a new loss and a new indicator to address the weakness of IoU. Based on this, this paper innovatively incorporated this Generalized IoU (GIoU) as a loss function into the most advanced SSD object detection network model, and carried out experiments on the original model and the improved model respectively based on the standard detection data set PASCAL VOC. The experimental results proved that the improved model had higher accuracy and better effect.