Optimizing the F-measure for Threshold-free Salient Object Detection
本文的出发点主要是基于CNN的显著性目标检测主要依赖于对交叉熵损失的优化。然后被检测的显著性图经常通过F-measure进行衡量。这篇文章调查了一个有趣的问题：在训练和衡量阶段能否一致使用F-measure？通过重新定义标准的F-measure，提出relaxed F-measure。与传统的交叉熵损失相比较，梯度在饱和区域降低更快，这个损失函数称为FLoss，甚至当激活接近目标时也有相当大梯度。因此，FLoss可以不断的迫使网络产生极化激活。
提出了FLoss有三个属性：
1） Threshold-free salient object detection. Models trained with FLoss produce contrastive saliency maps
in which the foreground and background are clearly separated. Therefore, FLoss can achieve high performance under a wide range of threshold.
2） Being able to deal with unbalanced data. Defined as the harmonic mean of precision and recall, the F-
measure is able to establish a balance between samples of different classes. We experimentally evidence
that our method can find a better compromise between precision and recall.
3） Fast convergence. Our method quickly learns to focus on salient object areas after only hundreds of iterations, showing fast convergence speed.
FLoss推导公式参考论文。