Physics Object Localization with Point Cloud Segmentation Networks

In modern particle physics experiments, the identification and trajectory of physics object, e.g. leptons and jets, depends on a complex pipeline of feature extraction, search, and machine learning algorithms. Deep neural networks (DNNs) offer a possibility of streamlining this process. This note describes the production of a dataset that the ATLAS collaboration is releasing for community use. It is formatted like popular point-cloud datasets, e.g. ShapeNet or S3DIS, which have been used to research new network topologies and applications. This dataset broadens the domain of these public datasets to include high energy physics detectors and encourages new deep learning research. The note also describes the application and performance of the PointNet++ and DGCNN networks, as well as the GravNet and GarNet networks developed by the HEP community. These networks predict to what type of physics object each detector readout channel is associated, e.g. electron, jet, or neither in this case. In the analysis, the performance is measured using the mean Intersection over Union scores. The result suggests that these networks could be used as a method to narrow the search space in a traditional reconstruction pipeline. The dataset is derived from simulated $Z \rightarrow e^+ e^-$ production in association with jets. Each event is represented as a list of non-zero ATLAS detector readouts, including the detector channel readout position, raw readout value, and some truth information. The truth information indicates if a channel is associated to an electron, jet, or neither. There is also tracking information included to facilitate track studies.