Hi all,
I’m using Pix4D Enterprise for drone surveys and feeding the per-point uncertainties from result_point_uncertainties.csv into a downstream probabilistic modelling pipeline.
Currently the CSV exports the marginal standard deviations (sx, sy, sz) per 3D point. However, from the Pix4D error estimation documentation (Strecha, 2014 — https://data.pix4d.com/misc/KB/documents/error_est.pdf), I understand that Pix4D internally computes the full 3×3 covariance matrix per point via S²₀(AᵀA)⁻¹, which includes off-diagonal terms capturing correlations between axes.
For my application, the off-diagonal terms matter — they affect directional uncertainty estimates when the viewing geometry creates correlated errors between axes (e.g., oblique views coupling X and Z uncertainty). Right now I’m using the diagonal as an approximation, which works but leaves accuracy on the table.
Would it be possible to add an export option for the full symmetric covariance matrix per point? That would be six values per point (σ²_xx, σ²_yy, σ²_zz, σ²_xy, σ²_xz, σ²_yz) — or equivalently, writing the upper triangle of (AᵀA)⁻¹ scaled by S²₀.
I imagine this would be useful to anyone doing uncertainty-aware downstream processing: probabilistic surface reconstruction, deformation monitoring, change detection, sensor fusion, etc.
Cheers!
Liam