Multi-body SE(3) Equivariance for Unsupervised Rigid Segmentation and Motion Estimation (Supplementary Material)

The prediction module of vanilla EPN is designed for global SE(3)-equivariance for all input points. For this purpose, we further devise two heads for rigid segmentation and motion estimation. Figure 1 demonstrates the detailed design of our EPN feature extractor on SAPIEN, OGC-DR, and OGC-DRSV, and that on KITTI-SF shares the same structure but has larger numbers of output dimensions accordingly. Figure 1: Structure of our feature extractor based on EPN. "EPNConv" is the SE(3)-equivariant convolution proposed in the vanilla EPN network.