Why don't you use IOU for training?

Question

When people try to solve the problem of semantic segmentation using CNN, they usually use softmax-crossentropy loss during training (see Fully conv. - Long ). But when it comes to comparing the performance of different approaches, measures such as cross-connect are reported.

My question is: why don't people train directly to the extent that they want to optimize? It seems strange for me to train to a certain extent during training, but to evaluate in a different way for tests.

I see that IOU has problems for training samples where there is no class (union = 0 and intersection = 0 => divide zero by zero). But when I can ensure that every sample of my truth is based on all classes, is there another reason not to use this measure?

Answer 1

Make this article where they come up with a way to make the IoU concept differentiable. I implemented my solution with amazing results!

Answer 2

: " -, ?". - , . (, , ). (softmax crossentropy) . , -, , , , , , - (, , , ).

Answer 3

.

, , ​​ IoU, . , , .

An assessment of this direction is that earlier comments mean that errors are differentiable . I believe that there is nothing about the IoU indicators that the network can use to say: "Hey, this is not quite here, but I should perhaps move my bounding box a little to the left!"

Just subtle information, but hope this helps.