Tree Segmentation

Goal

The goal of this project is to develop a method to segment and identify trees using remote sensing data. The two main types of data will be RGB images (or other wave lengths) and LiDAR point clouds. The main focus of the project will be to properly identify and separate each individual tree, even in dense forests, including covered trees which are trees invisible from above because of occlusion by a larger tree.

Setup

Requirements

Before running this repo, you need to make sure that you have CUDA installed if you want to use PyTorch with CUDA.

Also, the data that is used takes up to 20 GB of disk space, and the weights of each model take almost 150 MB. Therefore, you need to have enough disk space to run the project.

Set up the environment

First, clone the repository with its submodules:

git clone git@github.com:sogelink-research/tree-segmentation.git --recursive

Then, move into the repo folder and create the conda environment to run the project with conda:

cd tree-segmentation

conda env create -n <env-name> -f environment.yml

Execution

To run the training or the evaluation of a model, you simply have to modify and execute run.py. You can select the hyperparameters you want to use and launch one or multiple training sessions.

All necessary data (RGB images, CIR images and LiDAR point clouds) will be downloaded and pre-processed the first time you need it, and stored in the data folder.

Internship report

The report of this project can be found at this link: https://zokszy.github.io/Geodan-internship-report/