Frank Dellaert
Frank Dellaert
+1: Ran into same issue. While I could install libvips fine, `npm install numjs` still failed with ``` npm ERR! make: *** [Release/nothing.a] Error 1 npm ERR! gyp ERR! build...
I actually like suggestion (1) quite a bit, and in fact could that not be done *always* ?? You'd just have to make sure the default implementation does the right...
It's a bit hacky. I still think an API change that has an optional method to compute m*d jacobians would be the best way to go. It would indeed need...
> I suspect that suggestion 2 will give you a substantial performance boost. Not so much. 77% -> 71% of Evaluate. Looking into other thing...
Suggestion 1 is 20% *more* expensive. The special structure - and there is one - is mostly present in the m*D matrix, and is unavailable in MultiplyByJacobian. 't was a...
No autodiff :-)
I found a *bit* of special structure, but it does not make a big change. m*d is the way to go here (as in GTSAM :-)
Parting thought for today: things would already be much better if the *cost function* could implement the chain rule. For example, a `FrobeniusPrior` factor I made on SO(n) has the...
Sure. This is part of Shonan Averaging, of course :-) I can even add you to the evolving repo. Here is n=8: ``` Number of poses: 1661 Number of constraints:...
n=8 is a rare case. But If possible I'd like to "flatten the curve" in n :-)