Guidance on the training method(code) using the KITTI dataset

[99angpaka]

Hello, Sarlin.

I have a question regarding the training method.
In table 3 of the paper, it seems that (c) is trained only on the KITTI dataset.

I attemped to train using only KITTI dataset by modifying the 'orienternet.yaml' > default > data : mapiliary to kitti.
However, when I tried training the results show -inf.
So, can you provide a guidance on the training method using the KITTI dataset?

Additionally, considering the different input image sizes between Mapiliary and KITTI, is it BEV's shape(size) also needs to change?
I am curious if there is any code to adaptively adjust the BEV shape based on the image size or focal length

I appreciate your response. Thank you!