Reference

B station video

Loss function for semantic segmentation

(cross entropy loss function)

1. Pixel-wise cross-entropy
2. Also need to consider the problem of sample equalization (some)
   
   The cross-entropy loss function here adds the pos_weight parameter.The proportions of the foreground and background are different, making the importance of each pixel inconsistent.In general, we set this parameter according to the proportion of positive and negative examples.

Here will explain the next two-category cross-entropy loss function (remove pos_weight):
The smaller the loss function, the better

1. log function is as follows
   
   As can be seen from the imageWhen the X value is closer to 1, the absolute value of Y is smaller, and when the X value is closer to 0, the absolute value of Y is larger.

eg:
y_true: true sample label 0, 1
y_pred: sample predicted label 0, 1

It can be seen from this that the closer y_pred is to the true label y_true, the smaller its loss function is, and can distinguish foreground label 1 and background label 0.
We also need to pay attention to the loss of judging the true label as 1 and the predicted label when y_true log(y_pred), think about it, when the true label is 0, this part is directly 0, so the other part is judged trueLoss of label 0

(Focal loss)

Ideas: Difficulty of pixels: The more difficult the sample, the greater the reward.
Implementation: Based on the cross entropy loss function, increase the difference between the real labels 0 and 1 and the predicted labels 0 and 1 $\gamma$Power parameter.

Assumptions$\gamma$=2
The image of the quadratic square function is as follows:

We can see that the function's growth trend between [0, 1]is increasing, so when the difference between the real label y_true and the predicted label y_pred is very large, the difference is close to 1, and the value after the square term is also very large. When the difference is close to 1, the value after the square termIt is also very small, so for the samples that are easy to be divided, the loss incentive is very small, and for the samples that are difficult to be divided, the loss incentive is very large.This makes a distinction between difficult and easy samples (or pixels).

Weights in combined sample size$\alpha$ (the ratio of positive and negative samples mentioned above pos_weight), it is the complete Focal Loss function.

Evaluation Criteria

IoU: Intersection and Union Ratio (the intersection of the yellow parts in the figure/the union of the yellow parts)

MIoU (calculate the average of all categories, generally used as a segmentation task evaluation index)