search

nmndeep / Robust-Segmentation

[ECCV 2024] Towards Reliable Evaluation and Fast Training of Robust Semantic Segmentation Models
19 stars 2 forks source link

About the reproduction of SegPGD #5

Open momo1986 opened 1 month ago

momo1986 commented 1 month ago

Hi Croce and all,

A great work. Congratulations!

I want to reproduce the work of SegPGD.

The default value is:

delta: float = 0.5, aux_weights: list = [1, 0.4, 0.4]

I am not sure whether this parameter is the setting of your paper and original SegPGD paper.

Thanks & Regards!

Momo

nmndeep commented 1 month ago

Hi, If I understand your query correctly, then you would find the attack settings we use for SegPGD in Appendix B in our paper. Specifically, we use PGD with best iterate selection with the following step-sizes (8e-4, 9e-4, 1e-3, 2e-3, 3e-3, 5e-3, 6e-3) for these radii (0.25/255, 0.5/255, 1/255, 2/255, 4/255, 8/255, 12/255) resp.

We did not train any robust models for SegPGD and I think these were not available from the original SegPGD work.

Hope this helps, Naman

momo1986 commented 1 month ago

Thanks Naman, is the "delta" for Dice Loss set as 0.5 is OK or should we set it to a different value?