# Structure-aware completion of plant 3D LiDAR point clouds via a multi-resolution GAN-inversion network

## TL;DR

This paper introduces MRC-Net, an unsupervised deep learning method for improving the quality of 3D LiDAR point clouds used in agriculture and robotics.

## Contribution

The novel MRC-Net framework uses a GAN-inversion strategy with multi-resolution principles to complete 3D point clouds without supervision.

## Key findings

- MRC-Net achieves high accuracy on virtual datasets with a CD of 8.0 and F1 score of 91.3.
- On agricultural datasets, MRC-Net preserves object integrity with CD 3.3 and F1 97.3 for regular cartons.
- For complex simulated plants, MRC-Net maintains shape with CD 8.6 and F1 88.1.

## Abstract

Three-dimensional (3D) point clouds acquired by LiDAR are fundamental for applications such as autonomous navigation, mobile robotics, infrastructure inspection, and cultural-heritage documentation. However, environmental disturbances and sensor limitations often yield incomplete or noisy point clouds, degrading downstream performance. This study addresses robust, high-fidelity point cloud completion under such practical conditions.

We propose an unsupervised deep learning framework, Multi-Resolution Completion Net (MRC-Net), which builds on ShapeInversion by integrating a Generative Adversarial Network (GAN) inversion strategy with multi-resolution principles. The architecture comprises an encoder for feature extraction, a generator for completion, and a discriminator to assess geometric integrity and detail. Two key designs enable strong performance without supervision: (i) a multi-resolution degradation mechanism that guides reconstruction across coarse-to-fine scales, and (ii) a multi-scale discriminator that captures both global structure and local details.

Extensive experiments on multiple datasets demonstrate that MRC-Net achieves accuracy comparable to leading supervised approaches. On virtual datasets (e.g., CRN), MRC-Net attains an average Chamfer Distance (CD) of 8.0 and an F1 score of 91.3. On a custom dataset targeting agricultural scenarios, the model preserves object integrity across varying complexity: for regular cartons, it achieves CD 3.3 and F1 97.3; for structurally complex simulated plants, it maintains overall shape while delivering average CD 8.6 and F1 88.1.

These results indicate that MRC-Net advances unsupervised point cloud completion by balancing global shape consistency with fine-grained detail. The method provides a reliable data foundation for downstream tasks—including autonomous navigation, high-precision 3D modeling, and agricultural robotics—thereby contributing to improved data quality in precision-agriculture and related domains.