About the role of weakclip

[holmeslong]

The paper introduces that the role of weakclip is to optimize the CAM seed to obtain more accurate pseudo-labels, and then use DeepLabV3+ to obtain the final segmentation result and test the quality of the pseudo-labels. However, the label used in weakclip training is cam seed, which can also be regarded as a pseudo-label. In this case, the result obtained by weakclip is the final segmentation result. Is there any problem with my understanding?

[litingyidalaoban]

The paper introduces that the role of weakclip is to optimize the CAM seed to obtain more accurate pseudo-labels, and then use DeepLabV3+ to obtain the final segmentation result and test the quality of the pseudo-labels. However, the label used in weakclip training is cam seed, which can also be regarded as a pseudo-label. In this case, the result obtained by weakclip is the final segmentation result. Is there any problem with my understanding?

I confused when i read the paper ，oops，i have the same question，bro.
I argue that

1. use mct former's cam as gt
2. train a weakclip that can predict pseudo mask
3. use weakclip's pseudo mask to train Deeplab