Open natureyoo opened 2 years ago
Hi, I implemented the code for UDA settings and it works well. Thank you. But, I found only ResNet50 config and codes although the experiment results were on VGG16 and Res101 backbones. Could you release vgg16 and res101 config files? I tried training with vgg16 backbone, but it doesn't work well. Thank you
Thanks @xinw1012 for sharing this work and @natureyoo for asking this question
I also implemented the UDA settings as mentioned in the paper. However, my results(mAP50) when using Res101 backbone didn't work that well.
For CityScape to Foggy, my results for (src-only, rotation, oracle) was (25.89, 35.38, 49.79), while the results in the paper were (25.6, 41.5, 48.6) I have tried to recreate it according to the description alongside the training config mentioned in the implementation details.
@natureyoo, could you share the numbers you got using Res50 and VGG or even Res101? Looking forward to some help and the code release for the UDA settings, thanks!
I'm sorry for the late reply. I've got mAP@50 (24.45, 35.45) for (src-only, rotation) using Res50 and (27.7, 34.94) with Res101. The result for Res101 would not be the optimal one because I didn't fine-tune it at all, but there was no performance gain than Res50.
Thank you for your interesting work. It was very impressive to improve the robustness to domain shift through self-supervision.
In particular, I'm interested in conventional UDA settings (e.g. cityscapes -> foggy cityscapes, Sim10k -> cityscapes) that you applied rotation operation as a self-supervision task. Could you release the code for this, as well?