From Forest to Urban: Data Efficient Tree Segmentation with Self-Supervised Pretraining on Height-Based Voronoi Maps

Abstract:
We present a data-efficient approach for tree segmentation that leverages self-supervised pretraining on height-based Voronoi maps. Our method significantly improves performance across diverse environments—from forested regions to dense urban settings—while reducing reliance on large annotated datasets.

Paper:

From Forest to Urban: Data Efficient Tree Segmentation with Self-Supervised Pretraining on Height-Based Voronoi Maps
BMVC 2025 Workshop Proceedings
https://bmva-archive.org.uk/bmvc/2025/assets/workshops/MVEO/Paper_5/paper.pdf

Project & Code:
See the project page and code: https://github.com/Jonetz/TreeDetection

Key Contributions

• Introduces a self-supervised pretraining strategy tailored to height-based Voronoi representations.
• Demonstrates robustness and generalization on both forest and urban tree segmentation tasks.
• Reduces annotation requirements while maintaining competitive performance.

Authors

• Jonas Geiselhart — University of Stuttgart
• Luca Reichmann — University of Stuttgart
• Alina Roitberg — University of Hildesheim

Links

• Workshop Proceedings: https://bmvc2025.bmva.org/proceedings/workshop-proceedings/
• Workshop Main Page: https://mveo.github.io/index.html