YOLO的Loss中为什么要对不含目标的检测框设置较小权值?

在YOLO v1中,loss里对不含目标(no object)的检测框(bounding box)进行了权值 λnoobj 的设置(0.5),在原文中解释为

Also, in every image many grid cells do not contain any object. This pushes the “confidence” scores of those cells towards zero, often  overpowering the gradient from cells that do contain objects.  This can lead to model instability, causing training to diverge early on. 

To remedy this, we increase the loss from bounding box coordinate predictions and  decrease the loss from confidence predictions for boxes that don’t contain objects. We use two parameters, λcoord and λnoobj to accomplish this. We set λcoord = 5 and λnoobj = 0.5.

我们知道在训练过程中(loss逐渐降低),不含目标的检测框其置信度向0收敛,包含目标的检测框的置信度向1收敛。由于最终的98个检测框中,大部分均不含目标,所以这种noobj的检测框在loss中占的比重更大,会导致其对总体loss下降的贡献也越大,使得包含目标的检测框向1收敛(训练的一个重要目的)的速度较慢。这不是我们想要的结果,因此对其设置权值可以有效地抑制这种情况的发生。