search

microsoft / ProDA

Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021)
https://arxiv.org/abs/2101.10979
MIT License
286 stars 44 forks source link

this amazing performance improvments kill this area!!!!!!!!!! #7

Open jingzhengli opened 3 years ago

jingzhengli commented 3 years ago

this amazing performance improvements kill this area!!!!!!!!!!

panzhang0212 commented 3 years ago

I think there is still a large gap with full supervised learning.

  1. recent self-supervised learning could get comparable results with supervised learning, why UDA is still worse than full supervised learning.
  2. ProDA follows self-training. Another solution for UDA is gap minimization, which is more general and needs more works to improve.