GANET: GRAPH-AWARE NETWORK FOR POINT CLOUD COMPLETION WITH DISPLACEMENT-AWARE POINT AUGMENTOR

Abstract

Remarkably, real-world data (e.g., LiDAR-based point clouds) is commonly sparse, uneven, occluded, and truncated. The point cloud completion task draws due attention, which aims to predict a complete and accurate shape from its partial observation. However, existing methods commonly adopt PointNet or Point-Net++ to extract features of incomplete point clouds. In this paper, we propose an end-to-end Graph-Aware Network (GANet) to effectively learn from the contour information of the partial point clouds. Moreover, we design Displacements-Aware Augmentor (DPA) to upsample and refine coarse point clouds. With our graph-based feature extractors and Displacements-Aware Transformer, our DPA can precisely capture the geometric and structural features to refine the complete point clouds. Experiments on PCN and MVP datasets demonstrate that our GANet achieves state-of-the-art on the task of shape completion.

1. INTRODUCTION

The rapid development of 3D scanning devices (e.g.. LiDAR) has provided an unprecedented ability to capture point clouds from complex 3D scenes. However, due to limited resolution and occlusion issues, the scanned point clouds are sparse and incomplete, which is why various applications such as 3D detection (Zhang et al., 2020b) cannot take full advantage of them. PointNet (Qi et al., 2017a) has attracted great attention to the learning methods on raw point clouds. Inspired by PointNet, PCN (Yuan et al., 2018) introduces a coarse-to-fine fashion to learning-based shape completion. Based on this fashion, subsequent work (Yuan et al., 2018; Tchapmi et al., 2019; Wang et al., 2020a; Liu et al., 2020; Wang et al., 2020b; Pan et al., 2021) investigates how to optimize the refinement stage for more detailed results. For example, SnowflakeNet (Xiang et al., 2021a) proposes snowflake point deconvolution to progressively refine coarse point clouds. Although difficult to discern as a whole, most incomplete point clouds maintain roughly recognizable contours. This observation motivates us to propose Graph-Aware Network (GANet), a novel graph-based network for shape completion. Compared with MLP-based methods previous approaches, which rely heavily on inductive learning and may neglect shape awareness as mentioned in Liu et al. (2019b) , graph-based methods can extract shape information from the hints of geometric relation more effectively. An overview of our GANet is shown in Figure 1 . Specifically, we design a Multi-scale Edge Aggregator (MEA) to extract expressive features with rich geometric information. The MEA first applies a set abstraction proposed by PointNet++ (Qi et al., 2017b) to reduce the point number of input data. This operation avoids the effects of noise and repeated points as well as reduces the model's computation complexity. To learn from the outlines of the partial point clouds, we construct the local graphs based on the neighbors of the given centroids. Then we propose a novel scalable module, Local Edge Aggregator (LEA) to process the local graphs. This module weights the importance of the edges in the local graphs and then aggregates the features of the edges for the output centroid features. In addition, to capture both the local and global structures of the input point clouds, we introduce the philosophy of multi-scaling to our LEA. Furthermore, we design Displacement-Aware Point Augmentor (DPA), a novel upsampling module to refine the coarse output. We leverage a multi-stage strategy to stack DPA blocks. In particular, we use the LEA as the feature extractor in every DPA block. The LEA can capture the geometric

