[question] Using FeatureMatching MatchFromKnownPoses When are positions known precisely enough?

[yan892]

Describe the problem

When are camera positions really known exactly? The wiki for Using known camera positions mentions the setting FeatureMatching: FromKnownPoses

In this post it is said that it can be helpful if the positions are known good enough, but can affect the quality if not. But at what point are positions known good enough?

For example, if I calculate the camera positions for the mini6 Monstree dataset normally and then specify them as known and select MatchFromKnownPoses in FeatureMatching, I get a lot more points than before. And many do not match. So the positions are not known precisely enough, are they?

Default Photogrammetry Pipeline:

Using the calculated camera positions as known camera positions:

Or when I test my data set, for example. The camera is fixed and does not move, but the object rotates. (I know that this is not optimal for photogrammetry). If I now have the camera positions calculated and specify them as known in another run and have activated MatchFromKnownPoses in FeatureMatching, I get the following:

Default Photogrammetry Pipeline:

Using the calculated camera positions as known camera positions:

Dataset :

https://github.com/yan892/Datasets-Photogrammetry/tree/main/Dataset1

Desktop :

• OS: win 10
• Meshroom version:
  • Binary version 2023.2.0