Recurrent Slice Networks for 3D Segmentation on Point Clouds
share
In this paper, we present a conceptually simple and powerful framework, Recurrent Slice Network (RSNet), for 3D semantic segmentation on point clouds. Performing 3D segmentation on point clouds is computationally efficient. And it is free of the quantitation artifact problems which exists in other 3D data formats such as voxelized volumes and multi view renderings. However, existing point clouds based methods either do not model local dependencies or rely on heavy extra computations. In contrast, our RSNet is equipped with a lightweight local dependency module, which is a combination of a novel slice pooling layer, Recurrent Neural Network (RNN) layers, and a slice unpooling layer. The slice pooling layer is designed to project features of unordered points into an ordered sequence of feature vectors. Then, RNNs are applied to model dependencies for the sequence. We validate the importance of local contexts and the effectiveness of our RSNet on the S3DIS, ScanNet, and ShapeNet dataset. Without bells and whistles, RSNet surpasses all previous state-of-the-art methods on these benchmarks. Moreover, additional computation analysis demonstrates the efficiency of RSNet.
READ FULL TEXT