Hacker News Reader

Comment by rsp1984 | Hacker News Reader

rsp1984

I'm not an expert in Probabilistic Graphical Models but I do know factor graphs well. I've also read quite a bit of the recent SLAM and VO stuff that came out of Dellaert's group.

Here's the thing: Dellaert loves to write about SLAM from a viewpoint of probabilities and densities. I think he likes to see himself as a mathematician. However at the end of the day he's working with Gaussian noise assumptions, just like everyone else, and the solution is obtained by some form of Least Squares, again, like everyone else.

I would really like it if he just cut all the probabilistic thicket and went straight to the core of how his methods improve the SOTA just in terms of how the friggin Least Squares problem is set up and solved (e.g. like here [1]). But I guess that would probably take away most of the "magic".

[1] http://grail.cs.washington.edu/projects/mcba/pba.pdf

View on HN

The factor graph representation and analysis lends itself to simplified understanding and reasoning about the relationship between poses, measurements, loop closures, etc. in a way that is simple and intuitive. Dellaert then uses that intuitive structure to show how to setup the problem in a way that can be efficiently computed using linear algebra.

To use another example: Let's say you understand and prefer to use QR factorization for some linear algebra use case. That doesn't mean that Cholesky or SVD are somehow "worthless" tools -- they just provide other useful insights or applications.

I personally find Dellaert's decomposition of the problem useful. It doesn't hurt that he was one of the early SLAM "inventors" (along with the likes of Fox, Burguard, et al). And since the topic of the post is (literally) probabilistic graphical models, it seems extremely relevant to this thread.

(I read through the article you linked. I do not agree that it is an "easier" or "simpler" formulation.)

> Dellaert then uses that intuitive structure to show how to setup the problem in a way that can be efficiently computed using linear algebra.

But Dellaert didn't invent Factor Graphs, nor how to map a Factor Graphs to a Least Squares problem. Others did that way before him.

My criticism is directed towards his style of writing about SLAM in terms of probabilities and densities, which, IMO, is moot since at the end of the day he's just using Gaussians as everyone else and hence everything boils down to a Least Squares problem again. His probabilistic viewpoint is just a detour, a distraction at that point.

If he contributed a different kind of solver that could deal with arbitrary distributions then his probabilistic viewpoint would make a whole lot more sense but as long as it's just Gaussians I just don't see the point.

You're wrong. What you're suggesting is there is no merit in going for an O(n^2) sorting algorithm to O(nlogn) and then to an O(n^1.2534) sorting algorithm.

SLAM at a multi-city scale can use mathematical trickery to improve runtime. Sure, a new solver would help - but we don't need it as long as there is another trick up the mathematician's sleeve. Oh, and it's harder to come up with a new solver / technique.

You cannot be more correct in your criticism, it feels as if there are a group of researchers who are just too enamored by beauty of "Bayesian stats and graphical models" that finding better solution of problem has become a secondary task for them.

Plus a segment of academia loves this type of research.

> However at the end of the day he's working with Gaussian noise assumptions, just like everyone else, and the solution is obtained by some form of Least Squares, again, like everyone else.

Well you need some kind of model to optimize, and finding the perfect one is a vastly more difficult problem that probably begins to incorporate some stuff from information theory (which starts to become rather a different field).

I don't actually see the issue with researchers working on algorithms to solve the Gaussian noise versions of these problems — they can usually be extended to some other model relatively easily (for instance, rotation synchronization via Riemannian manifolds can incorporate a Pseudo-Huber loss function without much difficulty).

Also, what is SOTA?

State of the art?

> Also, what is SOTA? State Of The Art

Reply on news.ycombinator.com