Value of kernel threshold

[JaWeyl]

Hi all,

I've a very short question regarding the kernel threshold t used to calculate the variance of the gaussian function (Eq. 3).

In your paper, you mention that you set t equal to 0.9.
However, in the readme of this repo you define the argument for the training script as

kernel-threshold 0.5

So should we instead set this argument to 0.9 to reproduce the results?

I'd appreciate any comments on this.

[wolny]

Hi @JaWeyl,

based on the ablation studies (see appendix A.7) in the paper the value 0.9 worked best for the CVPPP dataset. However we found out later that this value might be too restrictive for datasets with variable object sizes (e.g. Cityscapes), where the value of 0.5 works better. I left the value of 0.5 in the repository, since it's more generic across datasets.
Our ablation study also shows that it's more important to tune this parameter in sparse positive-unlabeled settings (see e.g. Fig. 10 in A.7). In the fully supervised setting any value between 0.5-0.9 range seems to work well.