search

Jacobi93 / Alpha-IoU

[NeurIPS 2021] Alpha-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression
GNU General Public License v3.0
154 stars 21 forks source link

question #5

Closed weizhao-11 closed 2 years ago

weizhao-11 commented 2 years ago

Yolov5's original loss function is ciou. When a = 3, the accuracy of using a-ciou will decrease by 1.2%. Is it true that a-ciou is not as effective as ciou?

Jacobi93 commented 2 years ago

Thx so much for your interest. We obtain consistent and noticeable improvement when alpha is in [2, 3] of alpha-ciou, using YOLOv5s on both PASCAL VOC and COCO.

weizhao-11 commented 2 years ago

Is the improvement compared with yolov5s ciou？

Jacobi93 commented 2 years ago

Yep, that's right.

weizhao-11 commented 2 years ago

Because I didn't see the comparison with ciou in the paper, do I have any doubts, so I want to ask you. Thank you very much for your answer.

Jacobi93 commented 2 years ago

No worries.