difference between two type of depth

[Vincento-Wang]

gaussian mean of z is the depth of render depth, but structure from motion(sfm) estimated depth looks like to be the ray depth of gaussian mean vector length, is it right ?

I see the code of depth loss between gaussian target view depth and sfm projected depth, why not just use two context view depth and gassian rendered depth to calc depth loss.

[Gynjn]

So, as you said we can use two context view depth and rendered depth to calc depth loss. But in this setting, we also utilize target depth also, that is only acquired from rendered gaussian.

[Vincento-Wang]

So, as you said we can use two context view depth and rendered depth to calc depth loss. But in this setting, we also utilize target depth also, that is only acquired from rendered gaussian.

from above, pose means context to target transform. pose_rev means target to context, why piexl aligned gaussian use the pose_rev to project get the mean(x,y,depth) of gaussian,

is there one I misunderstood ?

[Vincento-Wang]

thank you firstly, another question,

computed_depth means camera view depth, projected_depth means projected target depth, why this two can calc loss , I am confused,

[Gynjn]

For the first question, the output of pose network is w2c camera from input1 to input2, so for gaussian adapter, we use pose_rev for c2w. (from target to each view, because we assume the target view as an identity camera parameter.)

[Gynjn]

For the second one, that loss is for the geometric consistency, for more detail, please refer to the paper.

[Vincento-Wang]

For the second one, that loss is for the geometric consistency, for more detail, please refer to the paper.

paper link cannot open, could you provide the paper title.

[Gynjn]

Sorry about that, the title of paper is "Unsupervised Scale-consistent Depth Learning from Video".