search

YanchaoYang / FDA

Fourier Domain Adaptation for Semantic Segmentation
491 stars 79 forks source link

Unable to reproduce the good results of Synthia to Cityscapes #22

Open KaiiZhang opened 3 years ago

KaiiZhang commented 3 years ago

Unable to reproduce good results of Synthia to Cityscapes. In LB=0.01, the miou is 44.09 before ssl. In LB=0.05, the miou is 43.3 before ssl. Can you please provide the hyperparameter settings on the Synthia dataset or your intermediate weight？Thanks a lot.

victorvini08 commented 3 years ago

Heyy man...We are also trying to reproduce the results but in GTA->Cityscapes the evaluated miou is mostly coming 1-2 percent less than the reported value. Did you try to eval on the gta dataset? Also how much difference did you get in the results? Thanks!

KaiiZhang commented 3 years ago

Heyy man...We are also trying to reproduce the results but in GTA->Cityscapes the evaluated miou is mostly coming 1-2 percent less than the reported value. Did you try to eval on the gta dataset? Also how much difference did you get in the results? Thanks!

Yes, you are right. We faced the same problem in GTA->Cityscapes. In GTA->Cityscapes, we have overcome this problem with some additional trikes, such as: bigger input size, bigger batch size or warm up Strategy. However, in Synthia, even after using the trikes mentioned earlier, the result still did not meet expectations. Hope this helps you. If you have any progress, please feel free to discuss. And I really hope that the author could release the intermediate weights after stage 1.

saurabhmishra608 commented 3 years ago

@Zachzhang97 how did u setup the synthia training as the classes are different for both the datasets(synthia and cityscapes)?

Anurag14 commented 3 years ago

@Zachzhang97 were you able to reproduce the FDA with Single Scale performance? For the VGG-16 backbone, the FDA with Single Scale numbers are very low. Below is the performance for LB-0.01.

@YanchaoYang could you release the intermediate weights for single scale performance, that would be of great help. Or maybe point me about any tricks you used / tell us what was FDA with single scale performance on VGG backbone that would be reallly helpful :D ===>road: 44.96 ===>sidewalk: 19.51 ===>building: 56.82 ===>wall: 6.53 ===>fence: 6.94 ===>pole: 9.01 ===>light: 5.36 ===>sign: 4.42 ===>vegetation: 64.69 ===>terrain: 8.23 ===>sky: 51.78 ===>person: 32.89 ===>rider: 1.32 ===>car: 47.04 ===>truck: 3.21 ===>bus: 1.79 ===>train: 0.01 ===>motocycle: 4.98 ===>bicycle: 0.43 ===> mIoU19: 19.47 ===> mIoU16: 22.4 ===> mIoU13: 25.85