search

Eli-YiLi / WSSS_MMSeg

Pseudo-mask Matters in Weakly-supervised Semantic Segmentation
Apache License 2.0
18 stars 3 forks source link

Question about COCO14 segmentation performence #7

Closed Unrealluver closed 2 years ago

Unrealluver commented 2 years ago

Excuse me, I still find the performence of coco is only 33.64, could you help me with some details? I run the config file pspnet_scalenet101_40kx32_coco.py with val_mini.txt split using 4 gpus. The detailed report is as below: 20220203_085809.log

Eli-YiLi commented 2 years ago

My setting about batch size is 2img32gpus. You need to extend the iteration and reduce lr in linear policy, if your batch size is small (24).

Unrealluver commented 2 years ago

Thank you for your help, I will try it again, and could you find that few classes in the report have 0 mIoU, 1, 80, 89, etc. Its wierd. Do you get the same result?

Eli-YiLi commented 2 years ago

You may firstly change the batch size and get similar results. Few or no classes are 0 mIoU.

ZechengLi19 commented 2 years ago

Thank you for your help, I will try it again, and could you find that few classes in the report have 0 mIoU, 1, 80, 89, etc. Its wierd. Do you get the same result?

Same question. Do you figure it out?

Eli-YiLi commented 2 years ago

Yes, I meet it too. But I didn't analyze the reason. You can check the existence/number/quality of these labels in pseudo-mask.

ZechengLi19 commented 2 years ago

Yes, I meet it too. But I didn't analyze the reason. You can check the existence/number/quality of these labels in pseudo-mask.

I have download the pseudo-mask you provided. And I visualize it. However, it don't high light the "person" class. Therefore, I just evaluate the performance of the "coco_scalenet.pth" on coco2017(download from baiduyun). And the log file is as follow.(ps: I know you do this work on coco2014, but I just evaluate it on coco2017, sorry).

test_out_release.log

ZechengLi19 commented 2 years ago

Yes, I meet it too. But I didn't analyze the reason. You can check the existence/number/quality of these labels in pseudo-mask.

I would appreciate it if you can check your own log file and share it with me.