search

feipanir / IntraDA

Unsupervised Intra-domain Adaptation for Semantic Segmentation through Self-Supervision (CVPR 2020 Oral)
https://arxiv.org/pdf/2004.07703.pdf
MIT License
271 stars 35 forks source link

IntraDA trained from scratch or from InterDA params? #4

Closed zhangyuygss closed 4 years ago

zhangyuygss commented 4 years ago

Hey, how do you initialize the model in the intra-da stage, train from scratch, or resume from where the inter-da stopped? It seems that resume from inter-da step is a more reasonable way. By the way, how does the quality of masks of easy samples influence the intra-da step? Will not-so-good masks lead to worse results?

feipanir commented 4 years ago

For GTA5->cityscapes experiment, we actually trained from scratch during intra-da stage. As we compare "training from scratch" and "resume from inter-da", there is not much difference in segmentation task. But it is possible to be different in other experiments. For the quality of masks of easy samples, it really affects the final results so we suggest you to get cleaner maps for easy sample.