Depth accuracy calculation of measurements

[John]

Hi there,
I am using the zed2i as an outdoor measurement tool. I saw the depth accuracy is about 2%-5% when it comes to less than 10m. But when I use it as the measurement tool outdoor, I found the accuracy is quite bad when the distance from the camera is further than 6m.

Case 1. Distance: 9.3m. Traffic sign ground truth: 2.5m. Calculate from depth map: 6.06m. (very bad)

Case 2. Distance: 8.3m. Traffic sign ground truth: 2.5m. Calculate from depth map: 2.43m. (Good)

How I calculate the measurement: Euclidean distance formula.

Parameters:
Neural depth mode.
3.8.2 SDK.
Resolution: both 2k and 1080 are the same.

Is this situation to be expected? How can I calculate the furthest working distance in this scenario?

Another screenshot is showing that the same height fence will give different height results.

Thanks for your suggestion.

[Myzhar]

Hi @John
how did you calculate the distance by using the Depth Map?
Can you write down the formula/code that you used?
What I see in the pictures that you attached is not coherent…

[John]

Thanks for your reply.

There are two points for every lines. Say A(x1,y1) and B(x2,y2). Both of them are selected by mouse.
This is the function that I get (X,Y,Z) by (x,y).

def get_value(x, y):
Z = npz[‘depth’][y][x]
X = (( x - npz[‘cx’]) * Z) / (npz[‘fx’])
Y = (( y - npz[‘cy’]) * Z) / (npz[‘fy’])
return [X,Y,Z]

npz is dumped from the SVO file by below code.

        zed.retrieve_measure(depth, sl.MEASURE.DEPTH)
        cx = zed.get_camera_information().calibration_parameters.left_cam.cx
        cy = zed.get_camera_information().calibration_parameters.left_cam.cy
        fx = zed.get_camera_information().calibration_parameters.left_cam.fx
        fy = zed.get_camera_information().calibration_parameters.left_cam.fy
        timestamp = zed.get_timestamp(sl.TIME_REFERENCE.CURRENT)
        np.savez('./temp.npz', depth = depth.get_data(), timestamp=str(timestamp), fx=fx,fy=fy, cx=cx,cy=cy)

After that, I just use the Euclidean distance formula like below:

distance = math.sqrt((X1 - X2) * (X1 - X2) +
                        (Y1 - Y2) * (Y1- Y2) +
                       (Z1 - Z2) * (Z1 - Z2))

The close and horizontal situation is not bad, but the far vertical one is very wrong. The border is about 6m.

My device: ZED 2i 2.1mm with polarizer.

I tried to generate the point cloud and open it by open3d, then, I found that the pole was not straight.

Distance from the camera: 8m

Distance from the camera: >8m

View from the original position.