is it possible to use Kimera-VIO as a odometry for Dense 3D reconstruction system?

[Mad-Thanos]

It seems that Kimera-VIO only produce an optimized pose for each keyframe, and only fuse the keyframes to Kimera-semantics module. So the map will not be as dense as general 3d reconstruction, right?

Is it possible to make Kimera-VIO to estimate a 3d pose for every consecutive frame?

Thanks!

[ToniRV]

Hi @Mad-Thanos, good point! Right now we integrate pointclouds even for non-keyframes but we use the ROS TF interpolation between keyframes, which is definitely sub-optimal. At some point I was sending pose estimates at IMU rate to ROS, but with the latest heavy refactor I haven't been able to re-implement that feature.

[therishidesai]

When you integrate every pointcloud does the pose get updated also? Right now, I use registerBackendOutputCallback to setup a callback that gets the pose after every keyframe but I want to get the pose after each camera frame.

[therishidesai]

https://github.com/MIT-SPARK/Kimera-VIO/blob/8012623e0780a9d59ab2e27bbed68b163a94d1b4/src/pipeline/Pipeline.cpp#L134 I see that this is where the check for if it is a key frame happens. If I just had the backend run for every frame how much would that ruin the accuracy?

[ToniRV]

The accuracy won't get much worst, but you may not achieve real-time performance.

[ToniRV]

You can try to set the param intra_keyframe_time to 0.0 in params/Euroc/FrontendParams.yaml to have a pose per frame. This effectively makes each frame a keyframe.

[therishidesai]

Thanks. Do you plan on ever releasing the version that publishes poses at IMU rate?

[marcusabate]

Currently there is no plan to reimplement the IMU rate pose estimation