Test in tunnel about kiss-icp

[nachovizzo]

Hello @Cristian-wp, thanks for the question. We did try in some private datasets, and it was working (at least not drastically failing). I was looking for a publicly available datasets with such characteristics.

First the answer and then some analysis.

Dataset

Corridor SLAM example(Indoor10 Sequence), from TIERS/tiers-lidars-dataset-enhanced. I hope this is as “featureless” as you imagine.

Results

kiss_icp_pipeline data_3th_floor_2022-06-07-18-47-20.bag --topic /os0_cloud_node/points --visualize

Produces the following result

Analysis

Although we are not used to think it this way… the word feature is somehow associated with something “static”, like 1 unique point-cloud. If you analyze this point cloud in isolation, you will find no features, but as we explain in our paper, we exploit a prediction model, we are working with not just point clouds but sensor frames, which are produced at 0.05 seconds. Therefore, the paradigm shifts a bit, since the featureless of a unique point cloud might not affect much the registration results as long as the prediction model is doing a "decent job". Of course, that if you create a completely synthetic datasets with exactly no features at all, the system might fail. But we developed this with robotics applications in mind.