3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation

Figure 1 : Given an input 3D point cloud, our Multi Proposal Aggregation network (3D-MPA) predicts point-accurate 3D semantic instances. We propose an object-centric approach which generates instance proposals followed by a graph convolutional network which enables higher-level interactions between adjacent proposals. Unlike previous methods, the final object instances are obtained by aggregating multiple proposals instead of pruning proposals using non-maximum-suppression.