# Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans

**Authors:** Romain Loiseau, Elliot Vincent, Mathieu Aubry, Loic Landrieu  
**arXiv:** [2304.09704](https://arxiv.org/abs/2304.09704)  
**Date:** 2023-04-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.09704v1/

## Key points

- The Learnable Earth Parser is an unsupervised method for parsing large 3D scans of real-world scenes into interpretable parts.
- The goal is to provide a practical tool for analyzing 3D scenes with unique characteristics in the context of aerial surveying and mapping, without relying on application-specific user annotations.
- The method is based on a probabilistic reconstruction model that decomposes an input 3D point cloud into a small set of learned prototypical shapes associated with laser reflectance and colorized based on aerial photography.
- A novel dataset of seven diverse aerial LiDAR scans covering over 7.7km2 and a total of 98 million 3D points was introduced to demonstrate the usefulness of the results.
- The model provides an interpretable reconstruction of complex scenes and leads to relevant instance and semantic segmentations.
- This approach offers significant advantages over existing approaches as it does not require any manual annotations making it practical and efficient for 3D scene analysis.
- Evaluation metrics show that this method outperforms state-of-the-art unsupervised methods in terms of decomposition accuracy while remaining visually interpretable.
- This study presents an innovative approach that has potential applications in various fields such as urban planning, environmental monitoring, disaster response management, among others.
- The code and dataset used in this research are available online for further exploration by interested parties.

## Summary

The Learnable Earth Parser is an unsupervised method for parsing large 3D scans of real-world scenes into interpretable parts. The goal of this approach is to provide a practical tool for analyzing 3D scenes with unique characteristics in the context of aerial surveying and mapping, without relying on application-specific user annotations. The method is based on a probabilistic reconstruction model that decomposes an input 3D point cloud into a small set of learned prototypical shapes. These prototypes are associated with their laser reflectance (intensity) and colorized based on asynchronous aerial photography. To demonstrate the usefulness of the results, the researchers introduced a novel dataset of seven diverse aerial LiDAR scans covering over 7.7km2 and a total of 98 million 3D points, with diverse content and complexity such as dense habitations, forests, or complex industrial facilities. The majority of these points are annotated with a coarse semantic label such as ground, building, or vegetation. The model provides an interpretable reconstruction of complex scenes and leads to relevant instance and semantic segmentations. The quality of the reconstruction is measured using symmetric Chamfer distance between the input and output point clouds while only taking the points' positions into account (not intensity). If the points in prototype point clouds are associated with a semantic class, labels can be propagated from the reconstruction to the input. The evaluation metrics show that this method outperforms state-of-the-art unsupervised methods in terms of decomposition accuracy while remaining visually interpretable. This approach offers significant advantages over existing approaches as it does not require any manual annotations making it practical and efficient for 3D scene analysis. Overall, this study presents an innovative approach that has potential applications in various fields such as urban planning, environmental monitoring, disaster response management, among others. The code and dataset used in this research are available online for further exploration by interested parties.