5. Semantic segmentation validation

To train the semantic segmentation network, run the following command:

python3 fcn-12.3.1.py --train

At every epoch, the validation is also executed to determine the best performing parameters. For semantic segmentation, two metrics can be used. The first is mean IoU. This is similar to the mean IoU in object detection in the previous chapter. The difference is that the IoU is computed between the ground truth segmentation mask and the predicted segmentation mask for each stuff category. This includes the background. The mean IoU is simply the average of all IoUs for the test dataset.

Figure 12.5.1 shows the performance of our semantic segmentation network using mIoU at every epoch. The maximum mIoU is 0.91. This is relatively high. However, our dataset only has four object categories:

Figure 12.5.1: Semantic segmentation performance during training using mIoU for the test dataset

The second metric is average pixel accuracy...