XinyiYing / LESPS

Repository for "Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision", CVPR, 2023
https://xinyiying.github.io/LESPS/
102 stars 4 forks source link

训练epochs数的最优选择 #12

Open cannonli7 opened 10 months ago

cannonli7 commented 10 months ago

作者你好,有个问题想请教。figure 5中显示了不同情况下IoU随epochs的变化曲线,请问我们该如何确定IoU最大时对应的epoch?

XinyiYing commented 9 months ago

无法直接确定哪个epoch下IoU最大。IoU最大的这个节点和很多因素(目标的尺寸、信噪比、形状,标签的数量、位置)都有关系,因此不能使用一个epoch达到所有目标都能够取得最大IoU。在这个前提下,才有了标签进化框架(label evolution)的出现,利用网络预测的中间结果更新标签,具体细节见论文《Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision》