Is the model DeepLab v3+ ? about cag_uda

[lxtGH]

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

[RogerZhangzz]

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.

[RogerZhangzz]

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]

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.

[super233]

Is there anything new?