Segmenter run my own dataset, mIoU is decreasing (binary segmentgation)

[rouchoo]

!

[rouchoo]

I don't know why segmenter runing on my own vein dataset, the eval result is very low and keep decreasing. Please help! thanks.

[MengzhangLI]

Hi, try to use binary segmentation with latest MMSegmentation. See PR here: #2016

[MeowZheng]

Can you provide more details, like the model and the config you used?

[rouchoo]

Thanks for your attention!

1. The training process is strange, when trainig, loss and decode.acc_seg is normal, but it is abnormal when eval,like fig.1 and fig.2:

2. Here are my homemade dataset configuration:

1. ../config/base/datsets/mydataset.py
   

2. ../mmseg/datsets/mydataset.py
   

3. my training model (swintransformer) configuration:
   

[rouchoo]

My dataset contains image and label.
image is 5125123;
label is 512*512 (single channel) and the pixel values are 0 and 1.

[BalasubramanyamEvani]

I'm facing a similar issue. Did you find any resolution @Lenleooo ?