Multi-Band Densification

From a quick check with Piero and Steven, I’ve learned that ODM extracts Tie Points from all bands, but only does Densification on the data set’s primary band.

In reading someone’s post on an Agricultural Crop Monitoring group, it appears that Pix4D’s Densification and Tie Point Extraction are both only single-band (Green), leading their data set’s DSM to miss some sensed features in the environment (crop rows).

Knowing that the Blue band is incredibly noisy, and we typically can expect some loss of clarity as we approach NIR, what does everyone think about a weighted approach to using the data from the other non-primary bands for Densification?

It might help us to reconstruct more of the sensed phenomena, across all bands, more in line with how we extract our Tie Points!

1 Like

I think the main problem is that you might increase noise across the DSM (I don’t know for sure, but it would be my guess) if you use all bands for densification.

But might be worth doing some experiments and reporting back the results (if anyone has time to do this).

1 Like

How would we be able to test this within ODM?
Parameters to tweak, or source code changes?

I think the appropriate approach is to transform the data to some multi-spectral equivalent of Intensity, hue, saturation, and use the intensity band for the matching. For multi-spectral data, the equivalent might be a principle components transformation using the first PC.

So, the workflow would be to do the following:

  • Calculate the 1st PC of the mulispectral dataset and transform/squeeze into an integerized raster that MVE can use
  • Run the 1st PC through MVE for depthmap calculations
  • Return to using the remaining dataset for the rest of the processing.

There may be transforms other than principle components that would work, but PC is easy, generic, and understandable.

2 Likes

That sounds pretty robust, as the PCA should filter out the least significant/contributing pixels, so one would hope meaningless noise along with them.

Yup, exactly.

1 Like

How does that differ from how Tie Points are currently extracted from all bands?

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.