add depth absolute/relative encoding

[Data-Iab]

Two methods added to Depth:

• encode_inverse : invert depth with given scale and shift.
• encode_absolute : undo encode_inverse changes.

One method added to AugTensor:

• to_squeezed_numpy: as its name indicates, converts to squeezed numpy.

[thibo73800]

Once this last comment is solved + the pr description updated, we're gonna be ready to merge. I'll also ask @jsalotti to review since it could happen to use theses methods soon.

[thibo73800]

@jsalotti Can you quickly review this merge request. It propose some method that you might use at some point.

[jsalotti]

• The transformations are not consistent. Calling encode_absolute after encode_inverse will not give you the original value. This is due to the order in which you apply substraction and division.
• encode_inverse does not seem consistent with the convention of this paper. They apply scale and shift directly on idepth, whereas you apply it on depth before inversion. Is it on purpose ?
• in encode_inverse, it seems that the depth is modified in place instead of creating a copy to return an inverse depth

[Data-Iab]

• You're right, it's fixed now
• the encode_inverse is nothing but this with custom scale and shift.
• issue : the is_inverse change in "depth" when encode_inverse is called.

[Data-Iab]

It seems that we had a problem with functions name. Calling DPT with invert=False means that predictions are inverted (which is not compatible with the attribute name. here invert depth is returned). And so we can't call our function encode_inverse because there will be more ambiguity.
Hence, I changed methods names so encode_absolute is called to transform depth prediction.

dpt = DPT(invert=False)
depth = dpt(frame)  # inverted predicted depth
depth = depth.encode_absolute(shift=shift, scale=scale) # absolute depth

incode_inverse is called just to undo encode_absolute changes

[thibo73800]

@Data-Iab Sounds good. I think there is still a last comment unresolved