search

zhaozhengChen / ReCAM

The official code of CVPR 2022 paper (Class Re-Activation Maps for Weakly-Supervised Semantic Segmentation).
143 stars 18 forks source link

Training stages #7

Closed lsa1997 closed 2 years ago

lsa1997 commented 2 years ago

Hi, thanks for this excellent work. I notice ReCAM adopts a two-stage training strategy that trains FC layer-1 first. What if training FC 1&2 together directly? Will it be harmful to the performance?

zhaozhengChen commented 2 years ago

Thanks for your interest in our work. Yes, it would be harmful. The class-specific feature will be noisy if we train FC1&FC2 together from scratch, and then FC2 may overfit to the noisy features.

lsa1997 commented 2 years ago

Thanks for your reply!