arXiv-2022
Gevorgyan Z. SIoU loss: More powerful learning for bounding box regression[J]. arXiv preprint arXiv:2205.12740, 2022.
文章目录
1 Background and Motivation
目标检测算法流程中,计算损失函数是一个很重要的环节
DIoU、CIoU positively affected both training process and the final results
【GIoU】《Generalized Intersection over Union:A Metric and A Loss for Bounding Box Regression》
【DIoU】《Distance-IoU Loss:Faster and Better Learning for Bounding Box Regression》
we believe there is yet a room for drastic improvements
作者在 CIoU loss function 的基础上(distance, the shape and the IoU),提出 SIoU loss function(引入 angle),效果进一步提升
2 Related Work
CIoU、DIoU、GIoU
3 Advantages / Contributions
- 设计提出 SCYLLA-IoU (SIoU) loss function
- COCO 数据集上验证其有效性
4 Method
(1)Angle cost
minimize the number of variables in distance-related “wondering”
α \alpha α 属于 0-45°之间, α \alpha α 越大,angle cost 越大
否者上面的公式中 α \alpha α 全部换成 β \beta β,最小化 β \beta β
(2)Distance cost
把 angle cost 引入到了 distance cost 中
ρ \rho ρ 是横纵中心点的相对偏移,
ρ \rho ρ 变大(中心点趋向远离), Δ \Delta Δ 也会越大
ρ \rho ρ 变小(中心点趋向重合), Δ \Delta Δ 也会越小
c w c_w c