Hello, thanks for sharing the code!
Can anyone tell what should be considered source of truth? I have found a few differences between the paper and both versions of the code, e.g.:
1) initialisation of control points: in C++ and in the paper, PCA on X_world is used, while in Matlab it is just a unit simplex;
2) as a second step, the paper considers four options depending on the dimensionality of the null space; C++ only has three options (I have not figured out whether N=1 or N=4 missing); in Matlab, for N=1, the algorithm is not matching the pairwise distances between the control points but rather distances to the centre.
I want to open source a PyTorch version but don’t want to implement and benchmark all versions, sou would be good to know which is the source of truth. Thanks!
Hello, thanks for sharing the code! Can anyone tell what should be considered source of truth? I have found a few differences between the paper and both versions of the code, e.g.: 1) initialisation of control points: in C++ and in the paper, PCA on X_world is used, while in Matlab it is just a unit simplex; 2) as a second step, the paper considers four options depending on the dimensionality of the null space; C++ only has three options (I have not figured out whether N=1 or N=4 missing); in Matlab, for N=1, the algorithm is not matching the pairwise distances between the control points but rather distances to the centre.
I want to open source a PyTorch version but don’t want to implement and benchmark all versions, sou would be good to know which is the source of truth. Thanks!