Accuracy
It has made a number of evaluation indicators to assess semantic segmentation accuracy of these indicators is usually a variant pixel level of accuracy and IoU. We report the most commonly used evaluation of semantic segmentation, the method used to measure the performance of pixel by pixel tagging class. For ease of explanation, we give the following numbers mean: We assume that a total of k + 1 type, (to L0 from of Lk, including an empty class or background), Pij is belonging to class i are assigned to the j-th class was the number of pixels. That is to say, on behalf of the pii is the number of positive cases correctly classified, similar to several other p-values.
Pixel accuracy ( PA ) : This is the simplest indicator, the number of pixels is only calculated correctly classified with the ratio of the number of all the pixels.
Average pixel accuracy ( the MPA ) : slightly modified version of the PA, the ratio of correct classification of pixels is calculated according to the class, the final averaged by the total number of classes.
And an average deposit ratio ( MIOU ): This is the standard for the evaluation of segmentation, which is the calculated intersection of two sets overlap ratio set thereto and, in this problem, which is calculated dividing the true prediction system and the split between the cross ratio. This ratio may be redefined as a true positive cases (intersection) divided by the total number of (including true positive cases, the error cases, positive errors and negative cases (and sets)). IoU class is calculated by averaging the last.
Frequency weighting and cross ratio ( FWIoU ): is an improvement of MIoU, for each category is weighted according to the importance, from the importance of the frequency of its occurrence.
For all the indicators above, MIoU is the most common one, because it is the most representative, and very simple. The researchers used a lot of competition and the metrics to report their results.