Point Cloud Instance Segmentation for Spinning Laser Sensors

In this paper we face the point cloud segmentation problem from a Deep Learning (DL) perspective.
We focus in spinning laser sensors and benefit from the structured nature of the data: since all the measurements can be projected into a 2D grid without loss of information, we solve the perception problem in such space and then, we exploit the range information for ensuring 3D accuracy.
This allows us to (1) directly use state-of-the-art models designed for visual information and (2) effectively address the main challenges of applying DL techniques to point clouds, i.e. lack of structure together with increased dimensionality.
To the best of our knowledge, this is the first work that treats laser sensors as cameras, facing the challenges of re-training visual models to work with structured 3D data.
We face the data-hunger of DL techniques by introducing a novel data mining pipeline that enables annotating 3D scans without human intervention. Together with this paper, we present a new public dataset with all the data collected for training and evaluating our approach.
As experimental results show, our approach outperforms current techniques in three main aspects: (1) performance, as we achieve higher precision, (2) completeness, as we natively provide segmentation masks and (3) inference time.