A Light-weight Approach for Automatically Reconstructing Large-scale Trees

Wenmeng Zhou, Yao Yu, Yu Zhou, Sidan DuAdvances 2013 · DOI

Abstract

Different from previous tree modeling approaches, our method is based on the idea of making tree reconstruction as quick as possible and simplifying the representation of final results while keeping the tree model visually acceptable. Each tree is represented by a Billboard model. We first get the shape mask of a tree by projecting LiDAR point cloud onto a 2D camera plane. Then we use the shape fitting method to obtain the corresponding rotation axes and bounding boxes for the main trunk and tree crown. We get the corresponding texture and correct the misalignment artifacts by texture completion. Finally, we rotate each textured polygon around the rotation axis to a certain degree. We demonstrate the effectiveness of our system with some LiDAR data sets and compare our tree modeling scheme with other state-of-the-art reconstruction algorithms to show its advantages in terms of speed and memory footprint.

Introduction

Modeling trees is a notorious challenge in computer graphics due to their geometric complexity and structural diversity. Traditional methods generally fall into two categories: procedural modeling, which is computationally expensive and difficult to align with real-world data, and direct reconstruction from laser scans, which produces massive, memory-intensive 3D meshes. These high-fidelity models often prove impractical for large-scale urban simulations, such as virtual cities or flight simulators, where memory and rendering speed are critical.

This paper proposes a middle ground: a fully automatic, fast, and memory-efficient pipeline that produces "visually acceptable" models suitable for massive environments. The core philosophy prioritizes visual plausibility and low resource consumption over exact geometric replication.

Methodology

The proposed pipeline transitions from raw sensor data to a simplified 3D representation through several key stages:

Overview of the tree reconstruction pipeline, from input LiDAR and images to the final billboard model.
  1. Semantic Segmentation: Individual trees are isolated from larger LiDAR point clouds, separating them from surrounding structures like buildings or terrain.
  2. Shape Fitting: Instead of complex branching meshes, the 3D point cloud is projected onto a 2D plane to create a "shape mask." The RANSAC algorithm identifies the rotation axis of the trunk and a vertical axis for the crown, allowing the model to account for slanted trees.
  3. The Billboard Model: The tree is represented by just six polygons—three for the trunk and three for the crown—intersecting at different angles to provide a 3D appearance from multiple viewpoints.
  4. Texture Completion: To handle occlusions and artifacts (like sky visible through leaves), a Gaussian Mixture Model (GMM) clusters pixels in the HSV (Hue, Saturation, Value) space to remove non-tree elements. A connectivity-based tiling algorithm then fills gaps to create a seamless alpha texture.

Results and Discussion

The effectiveness of the system was evaluated using real-world data from mobile mapping vehicles equipped with terrestrial LiDAR and panoramic cameras. The results show significant improvements in efficiency compared to state-of-the-art methods like those by Livny et al.

Visual progression of the reconstruction process: (a) original image, (b) point cloud, (c) shape mask, and (d) the final light-weight model.
  • Speed: The pipeline reconstructs a tree in less than 2 seconds, a 10x speedup over high-detail alternatives.
  • Memory Efficiency: The final model size is approximately 40kB, compared to 25MB for traditional high-detail meshes. This massive reduction allows for the population of entire urban corridors without overwhelming hardware resources.
  • Visual Quality: By treating the trunk and crown as separate entities and using HSV-based texture cleaning, the models maintain high visual realism despite their simplified geometry.

Conclusion

This light-weight approach provides a scalable solution for large-scale urban modeling. While the billboard representation may lack detail at extreme close-ups, its speed and memory advantages make it ideal for broad virtual environments. Future work will focus on improving model precision and developing tools to convert existing high-detail models into this efficient format.

Summary updated

Abstract

Different from previous tree modeling approaches, our method is based on the idea of making tree reconstruction as quick as possible and simplifying the representation of final results while keeping the tree model visually acceptable. Each tree is represented by Billboard model. We first get the shape mask of a tree by projecting LiDAR point cloud onto 2D camera plane. Then we use the shape fitting method to obtain the corresponding rotation axes and bounding boxes for the main trunk and tree crown. We get the corresponding texture and correct the misalignment artifacts by texture completion. Finally, we rotate each textured polygon around the rotation axis to a certain degree. We demonstrate the effectiveness of our system with some LiDAR data sets and compare our tree modeling scheme with other state-of-the-art reconstruction algorithms to show its advantages in terms of speed and memory footprint.
BibTeX
@inproceedings{zhou2013light,
  author = {Wenmeng Zhou and Yao Yu and Yu Zhou and Sidan Du},
  title = {{A Light-weight Approach for Automatically Reconstructing Large-scale Trees}},
  booktitle = {Advances},
  year = {2013},
  doi = {10.2991/eeic-13.2013.17},
}
Figure