2-classed segmentation, IoU and Acc are 0.0 and nan respectively

[lifan724]

I set the num_classes==2 to model a 2-classes problem。The "CrossEntropyLoss" and "DiceLoss" both got validatied results show below:
+------------+-------+-------+-------+--------+-----------+--------+
| Class | IoU | Acc | Dice | Fscore | Precision | Recall |
+------------+-------+-------+-------+--------+-----------+--------+
| background | 95.43 | 98.46 | 97.66 | 97.66 | 96.88 | 98.46 |
| human | 0.0 | nan | 0.0 | nan | 0.0 | nan |
+------------+-------+-------+-------+--------+-----------+--------+

where the input label of my dataset is a gray mask with 2 different values 0 and 255. The loss configs is "loss_decode=dict(
type='DiceLoss', loss_weight=1.0)" or "loss_decode=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0))"

[MengzhangLI]

Hi, if you only have one foreground class, I suggest you use binary segmentation with latest MMSegmentation version and check out PR here.