In Open3D library there is a function which calculates the information matrix, it uses 2 clouds, a transformation matrix (output of a registration algorithm) and a distance. I would like to understand the meaning of a information matrix in the context of 3D point cloud registration, for what is it used for?
I know how to calculate it, but just because I read the Wikipedia article. And I read some articles, but there's nothing to guide me.
The derivation of the information matrix comes from equation 8 in the paper:
Robust Reconstruction of Indoor Scenes, CVPR, 2015 by Choi et.
Basically it's used for computing the sum of squared distances between all correspondences, and the paper did an approximation to construct the informaiton matrix.
By the way, the approximation is also used as virtual pair approach in a 1996 paper:
Multiview Registration for Large Data Sets by Kari Pulli. Pay attention to figure 6 if you're interested.
As pointed by @Jing Zhao, the derivation comes from the Robust Reconstruction of Indoor Scenes, but yet a more detailed one can be found on Observability, Covariance and Uncertainty of ICP Scan Matching by Barczyk, Bonnabel and Goulette, especially for the transformation linearization parts.
