1、inconsistency between metric (dominated by SkewIoU) and regression loss (e.g. ln-norms)
摘自KFIoU:
因为目标检测最终的评价指标还是以iou为主嘛,然后以smmooth l1 loss为损失的话,它与iou的变化不一致就导致了评价度量和回归损失不一致。
2、Moreover, in five-parameter system, parameters i.e. angle, width, height and center point have different measurement units, and show rather different relations against the Intersection over Union (IoU) (see Fig. 5). Simply adding them up for inconsistent regression can hurt performance.This issue may be mitigated by the eight-parameter system which use the coordinates of corners with the same units.
简单来说,就是在五参数定义法中,每个参数都是独立的,相互没有耦合关系, 每个参数对iou的作用不一样且相互之间没有制约,因此会造成回归难以收敛。