Could you support recall, F-score and Precision

[zhongqiu1245]

Motivation
They are very useful for mmsegmentation.

Related resources
https://github.com/qubvel/segmentation_models.pytorch/blob/master/segmentation_models_pytorch/utils/metrics.py

Thank you!

[sshuair]

@xvjiarui I am trying to add new feature support Recall, Precision and F-score. The core code has been implemented.

But there is a problem that needs to be discussed:

currently, the mIoU and mDice only return two metrics, IoU Acc or Dice Acc.

For f-score, usually these are three metrics Recall, Precision and F-score appear at the same time. But the mmseg/datasets/custom.py evaluate method only handle two metrics.

Should we consider refactor the evaluate and the metrics.py method?

Here is the proposal:

1. change the metrics.py/mean_dice metrics.py/mean_iou and metrics.py/mean_fscore function return value from tuple to dict.
2. the returned dict contain the metrics name to generate the summary data and print log.

[MengzhangLI]

It is an excellent suggestion and great PR.

I will close it but feel free to re-open it anytime you want.

Best,