MAG-Comp: Memory-Augmented Geometry-Driven Amodal Point Cloud Completion for Occluded Industrial Objects

Read full text Summarise with VIRAT
Authors: Daniel Tang, Kenneth Walker
Cited by: 0
References: 8

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Point cloud completion is crucial for robotic tasks, especially with occluded and noisy industrial data. While two-dimensional image guidance has been traditional, pure point cloud methods increasingly achieve state-of-the-art results, making amodal completeness—recovering both visible and occluded parts—critical for robust interaction. Inspired by these insights, we propose MAG-Comp, a novel framework maximizing geometric information for amodal point cloud completion. MAG-Comp utilizes a Hierarchical Geometric Feature Encoder, a Class-Agnostic Geometric Memory Bank for shape priors, and a Dynamic Amodal Region Inference Module for explicit occluded geometry reconstruction. Experiments on ShapeNet-Amodal and an Industrial Bin-Picking Dataset confirm MAG-Comp's superior performance, achieving a Chamfer Distance of 1.58x10-3 and an Amodal IoU of 54.20% on ShapeNet-Amodal, consistently outperforming state-of-the-art methods. The framework demonstrates robustness to varying occlusion, strong generalization, and competitive inference efficiency, making it suitable for real-time industrial applications requiring precise amodal three-dimensional representations.

Fields of Study & Subjects

3D Shape Modeling and Analysis Robot Manipulation and Learning Advanced Neural Network Applications

Publication details

Journal / source
Preprints.org
Publication date
31 March 2026
Language
EN
Document type
Preprint