Questions about the experiment on Cityscapes dataset about pytorch-deeplab-xception

[jfzhang95]

Sorry, currently I only train this model in Pascal Voc 2012. My suggestions are as follows

1. Previous implemented deeplab v3+ has some errors, you can check whether
   the model output_stride is 16. If it's not, you can replace the old model with
   new implemented model.

2. Current deeplab v3+ pretrain weights are load from xception. But I think it is
   not the most suitable way to load pretrain model. You can try some other ways
   to load pretrain model.

3. You can apply data augmentation to prevent overfit.

[jfzhuang]

Thanks for your reply and suggestions! And I wonder if you reproduce the experimental result on the Pascal VOC dataset as published in the paper?

[jfzhang95]

Currently not. This program is still in progress. If I reproduce the results,
I will update it in this repo~