Aggregative Input Convolution for Large-Scale Point Cloud Semantic Segmentation

Abstract

We propose an efficient semantic segmentation method for a large-scale point cloud. Previous point-based semantic segmentation methods to large-scale point clouds have been difficult. This is because those methods infer semantic labels to all the points used for feature extraction, and large-scale point clouds easily exceed their capacity. To solve this problem, we propose a novel point-based approach that predicts class labels for a downsampled point cloud and expands the labels to the whole point cloud by nearest-neighbor interpolation. The key idea of our approach is to give local features derived from the whole point cloud to each sample point by the newly developed Aggregative Input Convolution (AIC) and convert those features into wider context features by a point-based model for small-scale point clouds. AIC was experimentally confirmed to improve semantic segmentation accuracy on a large-scale dataset.