jmfacil / camconvs Goto Github PK

Thank you for your work. It's been a great inspiration to me, and I have a puzzle.

There is only camera with same focal length in my dataset. Can cam-convs be used for training using virtual cameras generated by resize or crop?

I see that the results in the paper all used multiple cameras with different focal lengths, but in some case the camera parameters are fixed, and I'm not sure if using resize or crop to change camera intrinsic parameters will work.

I'm look forward to hearing from you!

Hello！I am a little confused about this paper, 1) Does this network need camera parameters when inferring? 2）How does this network handle pictures of multiple scales?

In your paper, what was the kernel size of your CAM-Convs? I assume 1x1, but maybe 3x3?

Hello, I really appreciate your work. Through cam-convs we can input the camera intrinsics parameters separately, which greatly improves the applicability of the depth estimation model for robotics domain.
But when I download the source, I didn't find thew whole backbone network and the complete training scripts.
Is it convenient for you to continue to open source later?
If it can be provided, it will really help me a lot.
Thank you in advance !

Hi, Sir:
After reading your paper, I have a confusion that whether the Cam-Conv module or the data augmentation helps improve the accuracy? Thx.