Panoptic segmentation of NOvA particle detections using a point set transformer

NOvA is a long-baseline neutrino oscillation experiment that detects neutrinos from the NuMI beam at Fermilab. The functionally identical near and far detectors consist of alternating horizontal and vertical planes of PVC cells filled with liquid scintillator that provide two sparse 2D views of particle interactions. During reconstruction, the tasks of clustering hits into reconstructed particles and classifying these particles has commonly been done using a mix of traditional clustering approaches, Convolutional neural networks (CNNs), and other machine learning methods that process the sparse images by using dense matrices. We propose a method that operates on the sparse matrices with an operation that mixes information from both views, allowing for information to propagate between both views and for hits to be matched between views, giving a greater context for the model to classify the particle type. Our model greatly reduces the memory usage of previous method, while achieving competitive scores in both object segmentation and semantic segmentation.