0
$\begingroup$

I am curious, why would you ever choose direct slam(photometric consistency, etc) over feature based slam? Particularly for dynamic scenes?

It seems you're always better off aligning fewer points which you have a high confidence in over a lot of points you have low confidence in and are prone to poor data association - am I wrong?

For dense reconstruction this can be done after bundle adjustment, etc.

$\endgroup$

1 Answer 1

1
$\begingroup$

You can take a look at this paper: Feature-based or Direct

The choice would depend on the application and I would prefer direct methos when we don’t have a good source of odometry.

For example, in a drone with monocamera, we might have GPS and IMU, but they are typically not accurate enough for SLAM. I this case, direct methos like DSO will have an initial estimate for depth and minimize the error over time, converging to a good solution over time.

I automotive for example, vehicle have a good source of odemetry to start with and it’s common to use the ground plane to get the scale of the motion. In that case, it’s more efficient to use feature-based methods. Specialty, because features can be used for other functions running in parallel like calibration or free space detection.

$\endgroup$
1
  • $\begingroup$ Thanks, I guess my issue is it seems you're trading an easy problem for a hard one - essentially attempting full 3d reconstruction of the scene even on areas with low texture. This is doubly true in a dynamic environment. $\endgroup$ Dec 4, 2022 at 2:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.