Research on Multi-modal Point Cloud Completion Task

Abstract

With the wide growth of 3D data applications, point cloud completion is particularly critical in the fields of autonomous driving and robot navigation. Aiming at the problem that point cloud data is easily affected by occlusion and noise, these papers propose is a completion method based on multi-modal fusion strategy, which combines 3D lidar and structured light scanning modality information to achieve more accurate point cloud completion. Based on the point cloud data, we deeply analyze the sparsity and unstructured characteristics of the point cloud, explore the progress of multi-modal representation learning, and effectively apply GAN and self-attention mechanism to the completion task. This paper designs a hybrid encoder-decoder network architecture and integrates a special multi-modal feature extraction module, which can effectively capture the supplementary information from different modalities and improve the feature expression ability with the help of the self-attention mechanism. The experimental results show that the proposed multi-modal point cloud completion method has better completion effects than the current SOTA model, especially in dealing with point cloud data with highly missing and complex scenes.