jxhuang0508 / caco Goto Github PK

Hello, thanks for your amazing work!

I wonder if the 'caco_stage1.pth' is the best result or the last result of warm-up stage (i.e. results of 120K iterations).

Thank you very much.

Thank you for your wonderful work and I'm wandering if you could publish more training information about object detection ? Thank you very much!

Hello, thank you for your great works.

I cannot reproduce your reported performance with the following command

Is the above result in the second row(CaCo-S: 46.8 / CaCo-T: 48.3 / CaCo: 49.2) the results of warm-up stage?

I also want to know training details.

Thank you very much.

Hi! I'm embarrassed to leave a lot of issues :)

I leave an issue because I have a question about your training procedure.

After CaCo warmup stage, you fine-tune the model with ProDA, CRS, or FDA methodology.
I wonder if the CaCo objective function is not used in the fine-tuning process.

Even when I look at the command lines for CaCo + ProDA, I cannot find where CaCo is used. (From warm-up stage to Stage 3)

Thanks a lot.

Hello, thanks for your great work!

It seems that the objective function in finetuning stage includes 1) segmentation loss in the target domain (Cross-entropy) 2) Adversarial learning 3) Contrastive learning loss.

Also, it seems that you only use the target data (Cityscapes) in the finetuning stage.

Why did you not use source data in the finetuning stage, and did the adversarial training work well despite the absence of the source dataset?

Thank you very much.

Thanks for sharing the code. Could you please provide the appendix of your paper?

Thank you for your excellent work
Can you provide the results of non distillation of ProDA+CaCO in two data sets？

Thanks for sharing, but I only find code for segmentation, is there any code for the image classification task reported in paper?