mmsegmentation-distiller icon indicating copy to clipboard operation
mmsegmentation-distiller copied to clipboard

Published 20 hours ago • verified publisher icon pppppM

Question about teacher model performance in the paper "channel wise distillation for semantic segmentation"

Open ofishhaha opened this issue 3 years ago • 1 comments

Hi,

I happen to found that the results of PSPnet-r101 on Pascal VOC（mIoU=78.52, mAcc=79.57）is exactly the same as the official model provided by mmsegmentation.

1、Have you tested the pretrained model? or directly use the accuracy on the webpage? or train it independently? I have this question because I also tested the pretrained model and haven't got the same result. I will appreciate it if you can explain it a little.

2、Maybe it has an error in the mAcc result. because in the mmsegmentation webpage, 79.57 is reported as mIoU(ms+flip) not mAcc. After my test, the mAcc is about 84.22% for the teacher model. Is it my misunderstanding?

Thank you for your efforts.

ofishhaha avatar May 08 '21 09:05 ofishhaha

Sorry for the typo. The weight of the teacher model is the mmseg trained one, thus, the test results should be exactly the same as the webpage.

irfanICMLL avatar Nov 12 '21 07:11 irfanICMLL