Tree Segmentation Project

This repository contains the code for the master thesis titled "Enhancing Tree Segmentation on Large Forest Point Clouds with Synthetic Data"

Data Generation

The synthetic forest generation workflow is available at:

https://gitlab-ce.gwdg.de/hpc-team-public/synforest

Data Preparation

To generate a dataset from a LAS point cloud, run generate_datasets.py under preprocess/.

Tree Segmentation

There is a directory for each of the four segmentation methods: watershed, Dalponte, AMS3D, and SGPN: cls_watershed/, cls_dalponte/, cls_ams3d/, and SGPN/, respectively. The watershed, Dalponte, and AMS3D methods are implemented in R, while SGPN is implemented in Python using PyTorch.

To train SGPN, run SPGN/train.py <config> where <config> is the basename of the config file under SGPN/config/.

Evaluation

Each directory has a test.py and metrics.py file to run the test on a dataset with or without visualization, respectively.

Pretrained Networks

The pretrained networks are available under SGPN/models/

Citation

If you used this repository in your research, please cite as follows:

@mastersthesis{Doosthosseini2023,
  author      = {Ali Doosthosseini},
  title       = {Enhancing Tree Segmentation in Large Forest Point Clouds with Synthetic Data},
  type        = {Master Thesis},
  pages       = {66},
  school      = {Georg August University of Göttingen},
  year        = {2023},
}