COCO-Stuff failure to reproduce results

[AlaaAnani]

Hello, could you please add more details regarding the COCO-Stuff dataset in the data/cocostuff directory?

I assume that you use COCO-Stuff-10k v.1.1, which if downloaded, only has .mat files for the annotations. I acquired the .png labels format by using the transformation found in the mmsegmentation repository. This automatically changes the label <original image name>.mat files to <original image name>_labelTrainIds.png. Nevertheless, this doesn't match the naming convention in the val.lst file. However, even if I account for the different naming convention, the mIoU and the pixel accuracy are far too low (< 1%), which clearly indicates that something is either wrong with the dataset or the model.

Based on that, could you please elaborate further on the dataset creation process you went through?

[ODN81HC]

Hi, not sure if you have overcome this problem when evaluating COCOStuff dataset with HRNet here. I have also got too low mIoU (<1%) and not sure what did I do wrong here?
My guess is the pretrained model or at least the model they gave is not correct but not sure.