CUDA device-side assertions about gradslam HOT 1 CLOSED

Thanks for the interest @DiTo97

The device-side assert most commonly occurs when there's pose / depthmap errors, which then cascades through the data association step. I would recommend visualizing just a few frames first to see how the 3D scene reconstruction looks like.

Since you are using RTABMap to compute global poses, it is likely that the frames around which this assertion is triggered have pose errors.

======

As an aside, we have been able to successfully interface (well, patch) RTABMap and gradslam for use in the concept-graphs project (where I see you mentioned this issue).

You might want to look at this GH repo to see how we did that:

1. Follow this "before building..." instruction from the README -- we made a quick-and-dirty edit to the RTABMap script to hardcode camera intrinsics.
2. We also have instructions to export poses, and have provided a python script to convert the exported rtabmap poses to gradslam dataset format (you will want to use the AzureKinectDataset or replicate this code to suit your directory structure).

Needless to say, the intrinsics, image dims etc are assumed to be correctly specified in the yaml config file that you'd pass in.

===================

Hope this helps! Closing for now, but reopen if the repo I pointed you to doesn't address your question.

from gradslam.