The OAK-D Pro PoE (Series 2) running DepthAI 2.28.0.0 is producing depth values that appear both incorrectly scaled and sparse when used indoors on a pickleball court.
Issue summary:
Depth scale appears off by roughly 4×.
Measured distances are around 3–4 ft where the real-world distance is 15 ft.
Depth map contains large invalid (black) regions, especially on smooth, painted court surfaces even with IR illumination enabled.
Both RGB and depth streams record correctly and align visually, but the absolute depth values are too small and the valid pixel count is low.
Test description:
Ten RGB frames and ten matching depth (.npy) frames were captured from a fixed indoor setup.
Two known ground points (court lines) that are exactly 15 ft apart consistently measure around 3.8 ft in both 3D and ground-plane distance.
This suggests a scaling or calibration issue related to stereo baseline or disparity-to-depth conversion.
Camera setup:
Device: OAK-D Pro PoE (Series 2)
DepthAI version: 2.28.0.0
Mount height: approximately 10 ft (7 ft tripod + bench), tilted about 45° downward
Environment: indoor pickleball court, even lighting
RGB stream: 1280×800 @ 30 FPS
Depth stream: 800p
Stereo settings: Subpixel=True, LR-Check=True, ExtendedDisparity=False
Median filter: 7×7
Confidence: 150
Alignment: stereo.setDepthAlign(dai.CameraBoardSocket.CAM_A)
IR illumination: Dot=900, Flood=900
Calibration: per-session calib.json (fx≈800, fy≈800)
Software: DepthAI SDK + Python 3.12 on Windows 10
Request for guidance:
Could you please review my recording script to confirm whether any configuration or parameter choices might be causing the incorrect depth scaling?
Could this issue stem from calibration — for example, an incorrect stereo baseline or disparity-to-depth conversion factor?
Would running Dynamic Calibration or re-flashing the factory calibration be advisable to verify the device’s accuracy?
Are there any example or reference pipelines that show correct real-world distances for comparison?
Any suggestions to reduce holes in the depth map on smooth indoor surfaces (e.g., adjusting IR power, confidence, or median settings)?
Depths: 2089 mm , 1764 mm
3-D distance = 1.312 m (4.30 ft)
Ground-plane distance = 1.189 m (3.90 ft)
Expected = 15 ft → Error ≈ 74%
Depth range = 0.63 – 65.5 m (valid ≈ 410k pixels)
Screenshot (wrong_length.png): shows 3.77 ft measured vs 15 ft actual.
Both RGB and depth data were captured using the same synchronized recording script.
Given the consistent scale compression and missing pixels, this appears to be a calibration or depth scaling issue.
Any confirmation, review, or working reference example that yields correct real-world distances would be greatly appreciated.
Data <-- dropbox Link to all the files i am talking about here.
Thank you.
Regards, Mohit
