I am using two ZED X cameras in a dual-stereo outdoor setup. Both sensors are fully calibrated, and I am running Neural Depth mode.
In deployments, I encounter major issues under strong sunlight (morning harsh reflections) and also when using a flashlight at night. Our target surface is highly reflective, which appears to severely degrade depth estimation.
The resulting problems:
-
One camera produces a systematically shifted point cloud (depth bias), while the other remains aligned → registration between the two fails
-
The degraded depth contains heavy noise artifacts
-
Large holes and invalid disparity appear in specular regions
-
The shift persists frame-to-frame, indicating biased disparity, not random noise
Calibration appears correct because the same setup performs well with controlled lighting.
Questions:
-
Is this a known failure mode of Neural Depth on reflective surfaces under strong illumination?
-
Are there recommended mitigation techniques (e.g., reduced exposure, depth mode changes, polarization filters, confidence-based masking)?
-
Any SDK settings to enforce more stability/consistency between the two stereo point clouds?
Any guidance would be greatly appreciated. I can provide images/logs if needed.