How to set depth_uncertainty when integrating different depth extractors

[Cyrus-Hao]

hi, We integrated PromptDA as the depth extractor in mpsfm.
Depths look accurate, but overall reconstruction is not ideal.
How do you adjust the hyperparameter "depth_uncertainty" when integrating other extractors?

[Zador-Pataki]

Hi,
Great to hear you’re testing new depth estimators!
MP-SfM has many components. What makes you think the uncertainties are the issue?
Calibrating them is tedious and requires calibrating using a dataset with ground-truth depth (see the paper for details).
I’d first confirm that uncertainties are really the cause before diving into calibration.

[Cyrus-Hao]

Thanks for the reply！☺️
We use PromptDA: absolute-scale depth (LiDAR-guided), but no per-pixel confidence. And we currently use a constant depth uncertainty and disable prior uncertainty.

Observation and intuition

• When dc check raises warnings, the overall score is strongly affected by t;
• t stability changes with depth_uncertainty: smaller values indicate greater trust in the model's depth;
• We therefore treat depth_uncertainty as a “global confidence” hyperparameter to better fit PromptDA and mitigate dc check warnings.

Question：
Is “dc check → score/t → depth_uncertainty” a reasonable causal chain? Any more reliable diagnostics or a better troubleshooting order? Or do we need to adjust the hyperparameters such as depth_cons_thresh to reduce the sensitivity of the dc check?

Thanks a bunch if you could tell me the parameter adjustment suggestions when there's no confidence map.

[Zador-Pataki]

Interesting. I'd be curious to see the failure cases.

MP-SfM has many mechanisms to handle unreliable priors. According to our observations, while monocular depth errors are only gaussian (in the log space) locally. This means that large portions of depth estimates need to be considered as outliers. In the BA, this was done by using robust losses, while in the DC-check, it was done by adjusting "t", to filter egregious errors. When LiDAR-guided, I would suspect that the distributions would better resemble gaussians, and so perhaps the "t" could be relaxed.

If I were you, I would visualize what kind of depth estimates are flagging the depth inconsistencies. As I don't have access to the data you are working with, it's hard to give you ideas about what might be happening, but yes what you explained makes sense. It's important to set the depth uncertainty correctly, so that the bundle adjustment is stable. Then, once that is settled, you can study the DC-check flags and adjust the hyper param. appropriately. If you don't need it, you can also disable it.