Abstract:The densification of a point cloud is a crucial challenge in visual applications, particularly when estimating a complete and dense point cloud from a local and incomplete one. This paper introduces a point cloud completion network named FuNet to address this issue. Current point cloud completion networks adopt various methodologies, including point-based processing and convolution-based processing. Unlike traditional shape completion approaches, FuNet combines point-based processing and convolution-based processing to extract their features, and fuses them through an attention module to generate a complete point cloud from 1024 points to 16,384 points. The experimental results show that when comparing the optimal completion networks, FuNet decreases the CD by 5.17% and increases the F-score by 4.75% on the ShapeNet dataset. In addition, FuNet achieves better results in most categories on a small sample dataset.

What problem does this paper attempt to address?

### Problems Addressed by the Paper This paper aims to address the key challenges of point cloud data in visual applications, particularly estimating a complete and dense point cloud from a locally incomplete point cloud. Specifically, the paper introduces a point cloud completion network named FuNet to tackle this issue. ### Background and Motivation As the most common 3D model representation format, point clouds have been widely used in fields such as computer vision and robotics. Due to their simple data structure and powerful expressive ability, point clouds play an important role in tasks such as 3D object classification, 3D scene segmentation, and 3D reconstruction. However, the point cloud data obtained in practice is often sparse and incomplete, which limits its direct application in some downstream tasks. Therefore, restoring a locally incomplete point cloud to a complete and dense point cloud is crucial for practical applications. ### Limitations of Existing Methods Existing point cloud completion networks usually adopt an encoder-decoder structure. The encoder is responsible for extracting point cloud features, while the decoder is responsible for generating a complete point cloud from a coarse point cloud. Existing methods are mainly divided into two categories: point-based processing and convolution-based processing. Point-based processing methods typically use multi-layer perceptrons (MLPs) to independently process each point, while convolution-based processing methods project the point cloud onto a regular structure for convolution processing. However, these methods usually consider only one processing method and cannot fully utilize the advantages of both methods. ### Design Concept of FuNet To overcome the above limitations, FuNet combines point-based processing and convolution-based processing to extract various features of the point cloud and fuse these features through an attention module to generate a complete point cloud. Specifically, FuNet achieves this through the following steps: 1. **Encoder**: - **Point-based processing**: Extracts local structural information through a multi-stage hierarchical structure. - **Convolution-based processing**: Extracts global contour information through 3D meshing. 2. **Decoder**: - **Feature fusion**: Fuses features extracted by the two processing methods through an attention module. - **Point cloud generation**: Generates a complete point cloud from the fused global features. ### Experimental Results Experimental results show that FuNet performs excellently in the point cloud completion task. On the ShapeNet dataset, FuNet's Chamfer Distance (CD) decreased by 5.17%, and the F-score increased by 4.75%. Additionally, FuNet also achieved better results in multiple categories on small sample datasets. ### Conclusion By combining point-based processing and convolution-based processing, FuNet effectively extracts various features of the point cloud and fuses these features through an attention module to generate a complete and dense point cloud. Experimental results validate FuNet's superior performance in the point cloud completion task.

FuNet: Multi-Feature Fusion for Point Cloud Completion Network