Timeline for How to optimize for the best fit nonuniform-scale-rotation to a given 3×3 matrix?
Current License: CC BY-SA 4.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 21, 2022 at 17:53 | comment | added | Alec Jacobson | Thanks. Don't I need to do svd each time I iterate on the sqrt? That seems to be the bulk of the Procrustes work so why not just update r? | |
| Nov 21, 2022 at 15:07 | comment | added | fedja | @AlecJacobson You iterate the square root until you arrive at the correct $S$ and then do one Procrustes to get $R$. The point is not that single iterations will get faster; they won't. However the square root method may need much fewer iterations than your current approach. | |
| Nov 18, 2022 at 21:21 | comment | added | Alec Jacobson | I'm a little confused how I would go about running these in parallel. It seems if I have to do SVD each iteration then the block-descent approach will always be strictly better. I'm probably misunderstanding. Here's what I'm trying gist.github.com/alecjacobson/70556ef7ed90a86f148404d167a2016a (gist so I can use newlines 😅). I think I'm confused about where I would be making the choice to use this $S=\sqrt{}$… versus $S = diag(diag(LR^T))$. | |
| Nov 17, 2022 at 22:43 | comment | added | fedja | Here is the promised link: mathoverflow.net/questions/413633/… As you can see, there seems to be no progress yet as far as the convergence is concerned :-( | |
| Nov 17, 2022 at 16:53 | history | answered | fedja | CC BY-SA 4.0 |