Improvement of projection-based LiDAR data segmentation algorithms using object-contextual representations

Nowadays self-driving cars and such unmanned aerial vehicles as drones are one of the most actively developed technologies, where machine learning algorithms play an irreplaceable role especially in the perception problem, which is the context of this research. To be applicable in self-driving cars and especially drones such algorithms should not only have good output quality, but also be real-time. For this reason, in the case of LiDAR data segmentation problem we pay special attention to algorithms that are based on point cloud projections because of their speed superiority over other heavy algorithms that process input point cloud directly. The main drawback of projection-based algorithms is their lower segmentation accuracy, so in this paper we show that it can be improved by integrating contextual representation module inside segmentation algorithm architecture. In our work we consider SalsaNext as a segmentation algorithm and OCR as a context representation module because these methods are among the highest in the corresponding datasets’ leaderboards. We provide results from quantitative evaluation on the Semantic-KITTI dataset, which demonstrate that the proposed SalsaNext modification gives 6.2% mean intersection over union metric improvement with no speed reduction.