ECG: Edge-aware Point Cloud Completion with Graph Convolution
Liang Pan
Abstract
Scanned 3D point clouds for real-world scenes often suffer from noise and incompletion. Observing that prior point cloud shape completion networks overlook local geometric features, we propose our ECG - an Edge-aware point cloud Completion network with Graph convolution, which facilitates fine-grained 3D point cloud shape generation with multi-scale edge features. Our ECG consists of two consecutive stages: 1) skeleton generation and 2) details refinement. Each stage is a generation sub-network conditioned on the input incomplete point cloud. The first stage generates coarse skeletons to facilitate capturing useful edge features against noisy measurements. Subsequently, we design a deep hierarchical encoder with graph convolution to propagate multi-scale edge features for local geometric details refinement. To preserve local geometrical details while upsampling, we propose the Edge-aware Feature Expansion (EFE) module to smoothly expand/upsample point features by emphasizing their local edges. Extensive experiments show that our ECG significantly outperforms previous state-of-the-art (SOTA) methods for point cloud completion.