x-axis coordinate warped shifts

[FrancsXiang]

x_t_flat = x_t_flat + tf.reshape(x_offset, [-1]) * _width_f

as i konw the offset should be pixel-level shifts for each pixel of x-axis,but what puzzles me is the meaning of ops "* width_f" back in the formula above.
could you explain it for any spare time ? :)

[jahaniam]

The way we are using it for every scale:
bilinear_sampler_1d_h(img, -disp) note that disp is not the offset disparity. It actually is offset disparity divided by the width of that in that scale to make it between 0-1.
That being said, to convert it to the normal offset pixel shifting known, you will need to multiply by the width.

disp = baseline * focal_length / (D * width) -> offset pixels = disp * width

Hope it helps.

[FrancsXiang]

Strike the key point! Thanks for answering my question.

disp = 0.54*tf.multiply(focal_length,invDepth)

Like you said before the disp is the normalized disparity.The baseline and depth should be in meters.
On the other side,the focal length should be in pixel_length,so i infer the disp should be the pixel shifts too.