Using custom images, how to deal with new camera calib?

[ghost]

Hey, thanks for sharing your amazing work!
I'd rather have a question: When I would like to apply your framework on custom images with camera intrinsics different to KITTI, do I have to retrain the models?
Or is there some kind of config-file? :)

Thanks alot!

[jake3991]

Any update on this? I'm also considering the same issue. Is there a recommended way to transform the depth image to a point cloud? How should I deal with the camera calibration?

Great work! This is a great repo!

[mrharicot]

@mojovski apologies for the massive delay.

Essentially the behaviour is undefined with images which have different aspect ratios and focals as the network really only saw one type of images.
Now with that in mind the best thing you can do is to convert the disparity into scaled depth by inverting it.

scaled_depth = 1 / pred_disp

[jake3991]

Ah I see. That being said you could set a camera to run a the required input resolution and re calibrate using for that resolution. Now you'd have a camera that can feed images to the network and convert disparity maps to point clouds. Do you think that is correct?

[mrharicot]

I am not sure I follow what you meant by re calibrating.
I think the most general case would require to train on a variety of aspect ratios and focals while keeping the total input resolution fairly constant.