Delayed depth initialization / Metrically-accurate maps with scale updates

[spillai]

I have a question regarding scale consistency/drift. I was wondering if the API permits intermittent scale updates (i.e. providing metrically accurate scale updates) to the pose graph in an on-line fashion. This would imply that providing a scale measurement at time t should technically propagate forward and backward in time over the pose graph, to be consistent with the metric map at any point of time (as long as there is at least one scale measurement available). I haven't spent too much time looking at the API to see if there's an obvious way of doing such an update. This would be particularly useful to obtain metrically accurate maps for future use.

Any help would be greatly appreciated. Thanks

[JakobEngel]

It is not directly supported by LSD-SLAM; However, it can easily be added into the pose-graph optimization back-end:

• If you know the scale of a specific keyframe (e.g. due to known depth / size of objects in it), add a unary edge to the pose-graph, which simply constrains the scale parameter of the keyframe's Sim(3)-pose (last value in log-representation).
• If you know the metric distance between two keyframes, add a binary edge with a scaled distance / pose constraint.

Doing so will implicitly propagate the absolute scale throughout the whole map.

[spillai]

Great, thanks for the response. Also, great work! :+1:

[icoderaven]

Hi

I've tried adding metric information to the pose graph using an external source of odometry for each keyframe by adding them as binary edges between the very first keyframe and the current keyframe in the findConstraintsForNewKeyFrames() method, and setting the threshold number of untracked keyframes after which to find new constraints to two. Although the pose graph updates the keyframes as I want to, what happens is that in the stereo image generation step, the keyframe poses used for the bunch of reference frames are the pre-graph optimized frames, leading to the point clouds not being registered with each other in the LSD SLAM viewer.

Is what I'm trying to do even possible? Because of the way the generated depth image depends on the tracking, it seems like it's non-trivial to get the latest keyframe pose for the stereo depth estimation.

Any help will be much appreciated! Thanks!

[icoderaven]

@JakobEngel Alternately, how do I add unary SE3 edges to the pose graph, such that the scale is propagated? Do I use a Sim3 datatype, with the information gain for scale set to 0, and that for the position and orientation pumped up? And what about the Jacobian in that case?

[ghost]

Hello @icoderaven. Were you able to find out how to add the unary SE3 edges to the pose graph? Thank you.