ForAINet

自动化森林清查：利用 3D 深度学习分析高密度机载 LiDAR 点云

该存储库代表题为“自动化森林清查：利用 3D 深度学习分析高密度机载 LiDAR 点云”的论文的官方代码。

设置环境

请参考我们之前的repo：

https://github.com/prs-eth/PanopticSegForLargeScalePointCloud

它包括详细的步骤以及可能发生但已解决的问题。

FOR实例数据集

请将旧的原始文件替换为我们的新原始文件：

例如，data_set1_5classes包含我们论文中表4中“基本设置”的数据。

1. 设置“基本设置”、“+二进制语义损失”、“+类权重”、“+高度权重”、“+区域权重”、“+弹性变形和子采样”、“+ TreeMix”的数据集

您可以下载它：

2. 如需其他设置，请在此处添加。

根据不同设置运行点云分割实验的命令：

 cd / $YOURPATH $/ForAINet/PointCloudSegmentation

1. 实验论文中的“基本设置”。

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1 models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1 job_name= # YOUR_JOB_NAME#

2. 论文中“+二元语义损失”设置的实验

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1 models=panoptic/FORpartseg_3heads_BiLoss model_name=PointGroup-PAPER training=treeins_set1_addBiLoss job_name= # YOUR_JOB_NAME#

3. 论文中“+类别权重”设置的实验

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_classweight models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1_nw8_classweight job_name= # YOUR_JOB_NAME#

4. 实验论文中的“+身高体重”设置

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_classweight models=panoptic/FORpartseg_3heads_heightweight model_name=PointGroup-PAPER training=treeins_set1_heightweight job_name= # YOUR_JOB_NAME#

5. 论文中“+区域权重”设置的实验

 # Command for training

# To be added

6. 实验论文中的“+强度”设置

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_add_intensity models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1_intensity job_name= # YOUR_JOB_NAME#

7. 论文中“+返回数”设置的实验

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_add_return_num models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1_return_num job_name= # YOUR_JOB_NAME#

8. 论文中“+扫描角度等级”设置的实验

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_add_scan_angle_rank models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1_scan_angle_rank job_name= # YOUR_JOB_NAME#

9. 论文中“+手工制作的特征”设置的实验

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_add_all_20010 models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1_addallFea_20010 job_name= # YOUR_JOB_NAME#

10. 论文中“+弹性畸变和子采样”设置的实验

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_curved_subsam models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1_addCurvedSubsample job_name= # YOUR_JOB_NAME#

11. 实验论文中的“+ TreeMix”设置

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_treemix3d models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=treeins_set1_mixtree job_name= # YOUR_JOB_NAME#

12. 不同点密度数据的实验

 # Command for training
python train.py task=panoptic data=panoptic/treeins_set1_treemix3d_pd#POINT_DENSITY# models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=mixtree_#POINT_DENSITY# job_name= # YOUR_JOB_NAME#

# take point density=10 as an example
python train.py task=panoptic data=panoptic/treeins_set1_treemix3d_pd10 models=panoptic/FORpartseg_3heads model_name=PointGroup-PAPER training=mixtree_10 job_name= # YOUR_JOB_NAME#

13. 用于测试的命令。请记住将“checkpoint_dir”参数更改为您的路径。

我们的预训练模型可以在这里下载：https://polybox.ethz.ch/index.php/s/aUypOXxOW3pYPe9

 # Command for test
# remember to change the following 2 parameters in eval.yaml:
# 1. "checkpoint_dir" to your log files path
# 2. "data" is the paths for your test files
python eval.py

# Command for output the final evaluation file
# replace parameter "test_sem_path" by your path
python evaluation_stats_FOR.py

运行树参数提取代码的命令：

 cd / $YOURPATH $/ForAINet/tree_metrics
# remember to adjust parameters based on your dataset
python measurement.py

用于运行提取手动提取的几何特征的代码的命令：

 # Please note that our code is based on the Superpoint Graphs repository, which can be found at https://github.com/loicland/superpoint_graph. We have included our custom partition_FORdata.py file.
cd / $YOURPATH $/ForAINet/superpoint_graph/partition
python partition_FORdata.py

引用

如果您发现我们的工作有用，请随时引用：

 @article{
  xiang2024automated,
  title={Automated forest inventory: analysis of high-density airborne LiDAR point clouds with 3D deep learning},
  author={Binbin Xiang, Maciej Wielgosz, Theodora Kontogianni, Torben Peters, Stefano Puliti, Rasmus Astrup, Konrad Schindler},
  journal={Remote Sensing of Environment},
  volume={305},
  pages={114078},
  year={2024},
  publisher={Elsevier}
}