SwiftPCN: Fast, implementation-efficient, and accurate point cloud completion network with flexible output resolution

Kevin Tirta Wijaya, Dong-Hee Paek, Seung-Hyun Kong

Abstract: Highlights•SwiftPCN: novel point cloud completion networks balancing accuracy and efficiency.•SwiftPCN-fixed: best speed-accuracy trade-off with prototypical encoder.•SwiftPCN-flex: variable output resolution via convolutional occupancy network.•Competitive accuracy to SoTA, 19x faster than SeedFormer, 20x fewer params than PoinTr.•Extensive experiments on synthetic and real-world data validate the superior performance.