About getting results in meters unit

[nightheronry]

@ranftlr Thank you for the work. I'm trying to apply it with Myriad X VPU.
So I would like to ask whether the unknown scale and shift mentioned in #36 are linear parameters?
For example, in each frame, I can find a linear equation like "P = D * scale + shift" to project the values of depth maps "D" to the physical absolute measurements "P" according to putting a known scale ruler in the view, right ?

[ranftlr]

Yes, they are linear parameters. When doing alignment, consider that P and D are inverse depth so you need to convert your absolute measurements to inverse depth first.

[nightheronry]

@ranftlr Thank you for your kind reply !! :)

[rodrigoGA]

If I understand correctly ...
P = is in inverse depth = 1/ physical distance in meters
D= inverse depth = result of the inference. It does not need inverse, the model already predicts the inverse
scale and shift = are estimated with least squares from known points in the image

Can I calibrate "scale and" shift "and use the model in an environment with the same lighting? Or is it necessary to calibrate in each frame?

[mahavir-GPI]

@rodrigoGA Did you get the correct scale and shift values based on the approach you mentioned?

[rodrigoGA]

according to my tests the model is very inaccurate to measure distance (uncalibrated in each frame ).
It's a shame, because it has very good runtimes and is compatible with tensorflow lite.
According to the internet there are approaches that take into account the previous frame that achieve better results

[mahavir-GPI]

Oh, To confirm, by uncalibrated do you mean you didn't calculate the scale and shift value? Did you directly use the model predicted values?

[puyiwen]

Yes, they are linear parameters. When doing alignment, consider that P and D are inverse depth so you need to convert your absolute measurements to inverse depth first.

Hi, sorry to bother you. I use depth camera D435i , but the depth map is bad. So I want to use MiDaS to improve the quality of the depth map. I write the scrips to align the MiDaS and absolute depth map.
I need to confirm something. The alignment formula is y = kx + b, y means invert absolute depth with meters, x means the prediction of MiDaS, k means scale and b means shift. Am I right?
The align code is:
realsense_depth_array_invert = 1 / realsense_depth_array # invert absolute depth with meters
x = midas_depth_array.copy().flatten() #Midas Depth
y = realsense_depth_array_invert.copy().flatten() # Realsense invert Depth
A = np.vstack([x, np.ones(len(x))]).T
s, t = np.linalg.lstsq(A, y, rcond=None)[0]
midas_depth_aligned_invert = midas_depth_array * s + t
midas_depth_aligned = 1 / midas_depth_aligned_invert
The midas_depth_aligned is finally absolute depth with meters. Am I right?
If I am right, I find aligned depth have large error. I dont know why, can you help me? Thank you very much!

[putrasto]

@rodrigoGA i think you need to calculate "scale" and "shift" for each image they can not be used for for other image because "D" is normalized.

[Shubhamkumarroy]

@ranftlr Thank you for the work. I'm trying to apply it with Myriad X VPU. So I would like to ask whether the unknown scale and shift mentioned in #36 are linear parameters? For example, in each frame, I can find a linear equation like "P = D * scale + shift" to project the values of depth maps "D" to the physical absolute measurements "P" according to putting a known scale ruler in the view, right ?

and can we use scale and shift ,same for all image if i have calculated using it two image if not then so for every image i have to calculate the scale and shift so how can it be don?