VV-Net: Voxel VAE Net with Group Convolutions for Point Cloud Segmentation
share
We present a novel algorithm for point cloud segmentation. Our approach uses a radial basis function (RBF) based variational auto-encoder (VAE) network with group convolutions defined on Z^3 to robustly segment point clouds without increasing the number of parameters. We transform unstructured point clouds into regular voxel grids and use subvoxels within each voxel to encode the local geometry using a VAE architecture. Traditionally, the voxel representation only comprises Boolean occupancy information, which fails to capture the sparsely distributed points within voxels in a compact manner. In order to handle sparse distribution of points within each voxel, we further employ RBF to compute a local, continuous representation within each subvoxel. We extend group equivariant convolutions to 3D point cloud processing and increase the expressive capacity of the neural network. The combination of RBF and VAE results in a good volumetric representation that can handle noisy point cloud datasets and is more robust for learning. We highlight the performance on standard benchmarks and compare with prior methods. In practice, our approach outperforms state-of-the-art segmentation algorithms on the ShapeNet and S3DIS datasets.
READ FULL TEXT