Mean IoU of 0.008, even when training to overfit.

[ghost]

I use my own dataset, but I train using just one image. The loss goes down to as low as 0.004. But when I evaluate using the same image I use for training, I get a mean IoU of 0.008. Why could this be the case? I made sure my image and label was being read correctly.

[Simpatech-app]

I have a same issue the loss is going to a small value,but the mean IOU is very bad.