stereo disparity vs depth map

[dimaxano]

As I understand you used stereo disparities for training, but in the paper, I see depth map should be used.

I see both are very similar modalities, does it really makes difference what to use as groundtruth: depth map or disparity?

[michael-fonder]

The depth maps you can download on the Mid-Air website are range maps. They give the Euclidean distance of an object to the camera.

In our paper (and in most depth estimation papers), we define the depth corresponding to a point as its "z" coordinate when its coordinates are expressed with respect to the camera frame of reference. This definition makes it different from the range map.

You could very well convert the range maps to "our" definition of depth by using trigonometry, but the conversion from the disparity map (which also gives a notion of depth) is easier mathematically speaking and less computationally expensive. So I computed the depth maps that fit our definition of depth by using disparity maps.