Update RAFT Stereo to be more sync with CREStereo implementation

[YosuaMichael]

After discussing with @TeodorPoncu for on the CREStereo implementation, we plan to make the implementation more consistent. Here are the changes in RAFT Stereo:

1. Change the input name from image1 and image2 to left_image and right_image
2. Enable using flow_init as input
3. Have self.output_channel as variable in RAFTStereo class

Here is the link to CREStereo implementation PR: #6310

[TeodorPoncu]

LGTM, just one question such that I know how to sync CREStereo to these changes.

[datumbox]

LGTM, I think the changes make sense and make the code much easier to read. Coming with basic knowledge of Depth Perception, the new naming allows me to understand what is happening and why.

[datumbox]

This is the only change for which I don't have context, but I understand it's a new feature.

[TeodorPoncu]

When wanting to perform inference at a resolution significantly larger than that at which the model is trained, you can perform cascaded inference.

Cascaded inference first computes the flow for a downsampled version of the image, and uses that flow as a prior for the full resolution image.

[YosuaMichael]

Just to give more context:
Here is the code for the cascaded inference: https://github.com/pytorch/vision/blob/test-crestereo-training/references/stereo_matching/evaluation.py#L57

And here is reference to original raft implementation on this part: https://github.com/princeton-vl/RAFT-Stereo/blob/main/core/raft_stereo.py#L104

[oke-aditya]

My doubt is do stereo models depend on order left and right input images. Can we not reverse them?
Like can we call the same model by changing left image to right and right image to left and still expect same output?

[YosuaMichael]

@oke-aditya no we can't reverse the input since what the model does is guessing how much "displacement" of the pixels from left_image to the right_image, and this usually mean all pixels will be shifted to the right (not the other way around).

We have however augmentation that do horizontal flip, and in this case we will also flip the order of right and left image.

[oke-aditya]

Perhaps we can document this.

[YosuaMichael]

@oke-aditya do you have any recommendation on where we should document it? I think this behavior will be general on all depth stereo model (currently we also plan to add CREStereo). Maybe we can add on the README when we add the references?

[oke-aditya]

Probably one root page for stereo models ? Also cc @datumbox for better ideas 💡

[YosuaMichael]

The test error is not related to this PR changes