search

prs-eth / DGInStyle

[ECCV 2024] DGInStyle: Domain-Generalizable Semantic Segmentation with Image Diffusion Models and Stylized Semantic Control
Apache License 2.0
63 stars 7 forks source link

Generated Datasets for Cityscapes #4

Open pmh9960 opened 2 months ago

pmh9960 commented 2 months ago

Dear Jia @yurujaja and Obukhov @toshas

I hope this message finds you well. Congratulations on your acceptance to ECCV 2024!

I am currently working on a domain generalization project on the Cityscapes -> ACDC, DZ, BDD, MV, I would like to compare my dataset generation method with yours. Could you kindly provide the generated dataset from Cityscapes that was used for the Table 2? This would be invaluable for my research.

Additionally, I am having difficulty reproducing the vanilla domain generalization performance of "DAFormer" and "HRDA" as presented in the same table. My reproduced results show that DAFormer performs worse (mIoU=50.0) without both SHADE and RCS as mentioned in this comment. However, using SHADE without RCS gives the best performance in my experiments (mIoU=54.27), which is close to the reported score. Can I use this score for my table, and could you please provide further details?

Model SHADE RCS mIoU
DAFormer X X 50.00
DAFormer X O 43.11
DAFormer O X 54.27
DAFormer O O 49.58

Thank you for your assistance.

Best regards, Minho Park

yurujaja commented 2 months ago

Thank you for your interest in our project. We will soon release the generated Cityscapes dataset. Please stay tuned for updates. Thank you!

pmh9960 commented 2 months ago

Thank you for your quick response and for sharing the wonderful work!

Siddharth-Shrivastava7 commented 3 weeks ago

Hi @yurujaja , Are there any updates on this?

Thanks !!