search

chenqi1126 / SIPE

[CVPR 2022] Self-supervised Image-specific Prototype Exploration for Weakly Supervised Semantic Segmentation
MIT License
73 stars 10 forks source link

The result of DenseCRF in Table 1 #8

Closed oooooppppsk closed 1 year ago

oooooppppsk commented 1 year ago

Hi, I try to reproduce the results of Table 1 in the paper. But the result of +DenseCRF is only 62.05 (using your trained weights to infer the cam file). Could you share your CRF process code and CRF hyper-parameters? Thanks!

chenqi1126 commented 1 year ago

Thanks for your attention. The CRF process is ensembled in training IRN. You can run cam2ir.py to obtain IR labels. Noted that the unconfident area is treated as background in evaluation for a fair comparison.

oooooppppsk commented 1 year ago

@chenqi1126 ,Thanks for your reply. So the result of +DenseCRF is the performance of IR Labels? And the evaluation file is eval_sem_seg.py? I am not sure whether my understanding is correct. Please correct me if I am wrong. Thank!

chenqi1126 commented 1 year ago

@oooooppppsk,

  1. If you read the cam2ir.py carefully, it is easy to find that the generation of IR labels is essentially a denseCRF process for the localization map, together with an auxiliary process for the background-weaken localization map that aims to determine the uncertainty area. When evaluating the IR labels, we treat the uncertainty area as the background. which equals the denseCRF process for localization map.
  2. eval_sem_seg.py is ok, but as I say, you need to add a preprocess to convert the uncertainty area (255) to the background (0) before calculation.
oooooppppsk commented 1 year ago

@chenqi1126 Thanks for your reply!