search

jiwoon-ahn / psa

Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation, CVPR 2018
MIT License
380 stars 62 forks source link

apply CRF during testing? #21

Open Juliachang opened 4 years ago

Juliachang commented 4 years ago

Hi @jiwoon-ahn,

I read your paper and you mentioned you applied CRF on CAMs and then use the CAMs to train the AffinityNet. My question is, did you also apply CRF on CAM during the test phase? Or you only applied CRF for training? Thank you.

Juliachang commented 4 years ago

Hi @jiwoon-ahn,

Sorry that I may misunderstand CRF part. I think you applied CRF on CAMs and then use that CAM to train AffinityNet. Then you use generated pseudo-groundtruthes to train the segmentation network. For testing, you only need to evaluate the output data from the segmentation network. Hope I understand correctly this time.

Here is another question, in your paper, Table1 and Table 5 show the mIoU of synthesized segmentation labels at different stages. I am wondering if the column CAM+RW+dCRF means the IoU of CAM--> apply dCRF --> apply random walk? If so, how did you apply random walk on the CAM that be refined by dCRF? Because the background is generated after applying dCRF, how did you handle background values when applying random walk?

I appreciate that you spend your time to reply this. Thank you.