ZED Mini depth scale issue

Hardware ZED Mini
1

Hi Stereolabs,

Our ZED Mini consistently overestimates distance in a simple tape-measure test:

1.00 → 1.03 (+3%)
2.00 → 2.12 (+6%)
3.00 → 3.34 (+11.3%)
4.00 → 4.66 (+16.5%)
5.00 → 5.76 (+15.2%)

Setup: HD720@30, depth mode NEURAL, depth_stabilization 70, min_depth 0.1 m, max_depth 15 m.

I’m a new forum user and can’t upload attachments yet, so I’m sharing view-only Google Drive links below.

Screenshots https://drive.google.com/file/d/12KIY-tJ7vfaIdA7p4_Ju5hDiRORlvGqv/view?usp=drive_link
Diagnostics JSON https://drive.google.com/file/d/1qDnoj9JZLHQgyfUL3zjrVFbDq9IzV6wa/view?usp=drive_link
YAML config https://drive.google.com/file/d/1OENnv5YGGXzfLOAqlhqguZB4PaXW_Kjv/view?usp=drive_link

Could you advise steps to fix this scale bias (recalibration, different depth mode, parameters to tune)?

Thanks,
Aidar

2

Hi @Aidar
Welcome to the Stereolabs community.

To check the behavior of your camera, please record an SVO in the same conditions using ZED Explorer and moving the target at different distances.

I recommend using a more visually informative target with texture. A white paper with text is not a good target for benchmarking.
Using an ArUco target would be the best.

3

Hi @Myzhar,
I’m working with @Aidar on this. We’ve recorded an SVO in the same setup using ZED Explorer and an ArUco target, sweeping distances up to 3m. Here’s the GDrive link: https://drive.google.com/file/d/1uzWlZJc5SxhOBmkjsZP7gULJ8fxkKUkh/view?usp=drive_link
Looking forward to your feedback. Thanks!

4

Hi @Taya
Welcome to the Stereolabs community.

This type of test requires a precise setup; otherwise, it’s not possible to validate the quality of the depth estimated by the camera.
You must use different targets at different fixed distances and measure the distance from the left eye of the camera to the target to obtain a valid ground truth.
Then you can validate the depth estimation by comparing what the SDK returns with the ground truth measurements.

Please use a rigid panel, non a flexible printed sheet.