Hi @shuguohui
It's simply because most of the time I observed that the ROTATION_AVERAGING_L2 method was giving me a better initial estimates for the rotation rather than the ROTATION_AVERAGING_L1 (I must admit that I still dislike that the L1 rotation averaging is starting for a MST). Ideally I would prefer having a global optimal minimization....

But since the rotation_averaging method is a command line parameter, feel free to choose the one that fit your needs.

On my side I prefer to consider that this problem "robust large scale rotation averaging" is not yet solved

• K. Wilson, D. Bindel, and N. Snavely, When is Rotations Averaging Hard? ECCV 2016. http://www.cs.cornell.edu/~bindel//blurbs/rotavg.html
• Anders Eriksson, Carl Olsson, Fredrik Kahl, Olof Enqvist, Tat-Jun Chin. Why Rotation Averaging is Easy 2017 https://arxiv.org/abs/1705.01362

ROTATION_AVERAGING_L2 can speed up by MKL,(when not use MKL,the cpu use only 10%,and use MKL,the cpu use more than 50%),in 10000 images test,use MKL is speed up 3-4 times

Using a Sparse Eigen value computation could made the thing faster and less memory hungry. Spectra library -> https://github.com/yixuan/spectra

@shuguohui Any help to port the L2 code to use Spectra?

Closed due to inactivity.
@shuguohui Feel free to reopen if you are willing to help