Using custom dataset

[poornimajd]

Hello great work @xy-guo and team!
I have stereo images and depth maps from zed camera for a custom dataset(realistic).In place of kitti scene flow dataset, I thought to initially train the stereo network on the zed camera depth maps.Then go for unsupervised stereo training on the same realistic data (with may be different scenes) and then do the final monocular training for the same dataset.
Will this result in improvements of depth estimation the way you have got in the proposed paper?Or basically will the final monocular depth estimation accuracy be limited to only the accuracy of zed camera depth maps?
Any suggestion is greatly appreciated.
Thanks

[xy-guo]

I recommend you to try recent stereo papers which focus on improving the generalization ability of stereo models, such as AdaStereo by Song et. al. Once you trained your stereo models following these papers, maybe you can directly get satisfactory results on realistic data. Actually I don't think zed can provide high-quality depth maps, since the depth is also computed by stereo algorithms? (I'm not sure about this). One suggestion is you can try to apply several different stereo algorithms to filter zed outputs to obtain some high-quality depth supervisions.
Steps:

1. train on scene flow using Adastereo (or some other stereo algorithms which solve the domain problem)
2. finetune on filtered zed data
3. mono training

[poornimajd]

Thanks for the quick reply!

Actually I don't think zed can provide high-quality depth maps, since the depth is also computed by stereo algorithms? (I'm not sure about this)

Yes this is true.

One suggestion is you can try to apply several different stereo algorithms to filter zed outputs to obtain some high-quality depth supervisions.

I will try out this,Thanks for the suggestions!