What is the future of filtering methods vs incremental-SFM in visual-SLAM

Question

In the Visual SLAM area, there's the well-known solution of EKF/UKF/Particle-SLAM , like the "mono-slam".

Recently , there was a direction to Local Bundle Adjustment methods , like lsd-slam or orb-slam ..

My question is :

Do the filtering ways still have a future or steady usage? in what applications? what are the pros/cons?

I read these papers but, I failed to extract a final answer,(mostly out of misunderstanding):

1. Visual SLAM: why filter?

2. Past, Present, and Future of Simultaneous Localization and Mapping

PS : I know the first is saying that Local BA is better somehow, and the second rarely mentioned filtering, so.., that's it.. , is it the end of the awesome Kalman filter in Visual-SLAM area?!!

Answer 1

No, the Kalman filter still has its uses. Although "visual SLAM: Why filter" is interesting in that it is the first (to my knowledge) paper to conduct a mathematically sound comparison, you should note that it only compares bundle adjustment to a very specific Kalman filter, which for example includes the points in the filter, while state of the art EKF-based odometry/slam methods seem to indicate that this is not a good idea. Also, you can argue that a recursive Kalman filter is more or less the same as bundle adjustment.

A kalman filter, despite its computational disadvantage in some cases, will have the advantage of easily providing you with uncertainty estimates. Obtaining non-local uncertainties in bundle adjustment is not trivial, and adds significant overhead (see for example this paper , which actually is the only uncertainty propagation in bundle adjustment I know of.).

Another advantage of Kalman filters is that sensor fusion is straightforward. You more or less have to add the parameters to estimate to the state vector. For an example of a state of the art Kalman filter for IMU/Vision fusion that is actually being used in many applications, see this paper .

But yes, there is a clear tendency in the SLAM community to move away from Kalman-based methods, except in specific areas (experimental sensors or large sensor graphs where having global covariances is mandatory etc), but the arguments are usually a little weak. People mumble something about better empirical results, and then cite "Visual SLAM: why filter". I recommend you read the thesis from that paper's author . Although his theoretical arguments on entropy are convincing, I still think that we have to be very cautious when citing that paper, because of the aforementioned particularities of the filter.

Answer 2

No, the second paper does not describe the end of the Kalman filter in Visual-Slam. The Kalman filter is a special case of the Maximum Likelihood Estimator for Gaussian noise. I want to direct your attention to Page 4, paragraph 3 of the second paper. There, the authors should clarify that the Kalman Filter and MAP are both extensions of Maximum Likelihood Estimation. As written, that insight is only implicit.