Point Cloud Transformer for Elementary Particle Signals Segmentation

Each experimental setup in high-energy physics experiments has its own specifics of tracking detectors and data acquisition system. For instance, SPD NICA track detectors will produce a huge number of fake measurements and other noisy signals, which can exceed the number of true ones by two orders of magnitude. In this paper, we present the Transformer-based architecture for the elimination of fake measurements from the simulated data for the SPD experiment. We describe an efficient method for utilizing self-attention modules with squared algorithmic and memory complexity to the simulated data by voxelization procedure.