SpadeLiu

> Hi @SpadeLiu > > I actually explicilty did this because if neither condition is met, there are NANs in the query and key. You can find more details in...

> I think the author put the test result(40 images) on KITTI2015 using the model trained on the training set(160 images). > But the final model put on Google drive...

I have the same question. The density of KITTI raw data is about 5%, while that of depth completion dataset is about 30%. Whether it will improve a lot if...

Thanks for your code. I also found you removed SPP module in your feature extraction module. To be honest, I think it is useless for matching especially those pooling layers...

Hi, the model submitted to the KITTI benchmark is trained with all the KITTI data (both 2012 and 2015), meaning that there are 394 training image pairs (194 + 200)....

> Hello, thank you very much for your code, I would like to ask why you need to input the , gt_tensor to predict the display when testing? > >...

> could you please offer the lac + ganet code and weights? Hi， For GANet, you have to compile first, and thus we don't provide the code. The most important...

> Hello, your Pretrained Models link is not opening, can you check it? It is ok for me. Could you please try with another browser?

Hi, We didn't do that since in this domain generalized problem we can use the training set to evaluate the model performance. Our method, unlike STTR or CFNet, is designed...

Hi, the results come from "Domain-invariant Stereo Matching Networks". And in "Matching-space Stereo Networks for Cross-domain Generalization", the authors use the same results.