HSG-Net: Point Cloud Completion via Heuristic Structure Growing.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Current point cloud completion methods struggle with complex partial data, leading to unsatisfactory results, especially in distant areas.
• Extracting latent codes from partial point clouds is insufficient for accurate reconstruction.

Purpose of the Study:

• To develop an advanced point cloud completion network that addresses limitations in reconstructing complex and long-distance structures.
• To improve the accuracy and quality of completed point clouds.

Main Methods:

• Proposed a novel heuristic structure growing network (HSG-Net) employing an iterative growth strategy for progressive completion.
• Introduced a data preprocessing (DP) method for ground truth (GT) generation with structural integrity.
• Utilized a consistency constraint displacement module (CCDM) and a feature memory module (FMM) for enhanced structure growth and quality.
• Incorporated a local information generator for final refinement of the completed point cloud.

Main Results:

• HSG-Net demonstrates superior performance in point cloud completion compared to state-of-the-art methods.
• The heuristic structure growth strategy effectively reconstructs close-distance details.
• The proposed modules (DP, CCDM, FMM) significantly enhance the quality of the grown structures.

Conclusions:

• HSG-Net offers a robust solution for point cloud completion, particularly for complex structures and areas distant from the input data.
• The novel approach achieves state-of-the-art results, advancing the field of 3D data reconstruction.