search

RogerZhangzz / CAG_UDA

(NeurIPS2019) Category Anchor-Guided Unsupervised Domain Adaptation for Semantic Segmentation
MIT License
141 stars 20 forks source link

Is the model DeepLab v3+ ? #6

Open river-afk opened 4 years ago

river-afk commented 4 years ago

Hello authors, thanks for the nice work and the published codes. I have a question regarding your segmentation framework. As stated in your paper, it is DeepLab v2. However when I look at your implementation of the ASPP module and the decoder, it seems to be the DeepLabv3+ (which performs better than DeepLab v2 generally https://arxiv.org/pdf/1802.02611.pdf ). Could you please confirm this point?

RogerZhangzz commented 4 years ago

Hi there. Yes it's DeepLab v3+. Very thanks for your pointing out and sorry for the mistakes. I would revise the paper. I would also re-implement the CAG model of the deeplab v2 version and report the experimental results if possible. Thanks again!

river-afk commented 4 years ago

Hello @RogerZhangzz , thanks for the prompt reply. In this case, to me it seems incomparable between your method and others. Basically using a more advanced framework like DeepLab v3+ will boost performance, no matter what UDA technique is used. I'm then more interested in results using DeepLab v2.

RogerZhangzz commented 4 years ago

Hi. Yes a more advanced model will improve the model. However, I re-implemented the baseline model based on Deeplab v3+. Then the performance sees an increase after CAG training, proving the effectiveness. Secondly, I also re-implemented the counterpart method of CAG, namely, self-training regarding the prediction probabilities based on Deeplab V3+. The results also improve by a large margin. Thirdly, the CAG method is compatible with most of other SOTA ones, and our implementation doesn’t incorporate them. Thus a further significant increase of performance can be expected when involving all other methods. Lastly, all ablation studies are based on Deeplab V3+, eliminating the possibility of performance improvements due to a better model. Thus the effectiveness of it has been validated in ablation studies. Besides, I will run the CAG of the Deeplab V2 version ASAP. Thanks.

lxtGH commented 4 years ago

Hi! what are the real results of Deeplabv2 in the paper? There is a huge gap between v2 and v3+ in performance. @RogerZhangzz

super233 commented 3 years ago

Is there anything new?