Uncertainty maps for facade parsing training

[manuelaceron]

Hello,
Great work!

I'm trying to replicate your results, and I have couple questions:

1. "train_facade_parsing.py" reads directly the image, ground truth and uncertainty map, how do you generate this uncertainty map? Do you have some inference script to generate it for the whole training/validation set using the model trained in "train_uncertainty.py"?

2. In the paper you mention 5-fold cross validation is used, however in the training procedure I don't find this is done, am I missing something?

Thanks!!

[wohaiyo]

We have two steps to train the proposed model. First, we train the Bayesian CNN model, we use the Bayesian model to generate uncertainty maps using in script "train_facade_parsing.py". The uncertainty map inference script is not in the repository. It is easy to write a script to realize.

You can write a script to implement 5-fold cross-validation.

[Abbsalehi]

Hi, nice work! I'm also replicating your research and have a question. Is there a specific reason why the Bayesian CNN model hasn't been validated to prevent overfitting? I noticed that the default training steps are set to 6000, which might result in the model overfitting. What are your thoughts on this? Thanks!