3D Point Cloud Paired-Attention Central Axis Aggregation Network

Network

Environment

• Ubuntu 22.04
• Python 3.8
• Pytorch 2.1.0

sudo apt-get install libsparsehash-dev

conda env create -f environment.yaml 

cd libs/pointgroup_ops
python setup.py install
cd ../..


# PTv1 & PTv2 or precise eval
cd libs/pointops
# usual
python setup.py install


cd libs/pointgroup_ops
python setup.py install
cd ../..


# PTv1 & PTv2 or precise eval
cd libs/pointops
# usual
python setup.py install


# Open3D (visualization, optional)
pip install open3d

FlashAttention

pip install flash-attn --no-build-isolation

Dataset Prepare

Simulation Method of Point Cloud Data for Maize Populations：

1. Download: Physically Based Deformation of Single Maize Point Cloud Datasets
   link
2. run

python project/multi_gen_group_data_no_land.py

Ground Truth Dataset

We conducted tests on a total of 17 datasets obtained from four types of sensors. The data catalog and test results are as follows:

Data ID	Data Name	AP
$A^1$	lidar__a.txt	0.7340
$A^2$	lidar__b.txt	0.7948
$A^3$	lidar__c.txt	0.8495
$A^4$	lidar__d.txt	0.9037
$B^1$	other__Maize-04_gt.txt	0.9808
$B^2$	other__grou_maize_gd.txt	1.0
$C^1$	slam__slam_all.txt	0.9868
$D^1$	rgb__0707_Tian_30_gt.txt	1.0
$D^2$	rgb__0707_502_30_gt.txt	1.0
$D^3$	rgb__0709_XY_20_gt.txt	0.8367
$D^4$	rgb__0709_XY_30_gt.txt	1.0
$D^5$	rgb__0721_Tian_20_gt.txt	1.0
$D^6$	rgb__0729_Tian_30_gt.txt	0.9738
	Average	0.9246
$E^1$	DjiV4_clean_gt.txt	0.9011
$E^2$	StPaulV3_clean.txt	0.9675
$E^3$	StPaulV6_clean.txt	0.5403
$E^4$	WasecaV5_clean.txt	0.6561
$A^1_{test}$	2-lidar__a.txt	0.7549
$A^2_{test}$	2-lidar__b.txt	0.7822
$A^3_{test}$	2-lidar__c.txt	0.8396
$A^4_{test}$	2-lidar__d.txt	0.9203

The ground truth of the test data and prediction results are published at the following address:

• datasets: link
• pred label: link

Additionally, we express our gratitude to several scholars who shared their data with us. We processed and annotated these data for testing purposes. The original links to these data include:

• other__grou_maize_gd
• other__Maize-04_gt
• uav__*

Train

python tools/train.py --config-file configs/corn3d_group/insseg-pointgroup-v2m1-0-pt3m2-base.py

Test

1. Change the configs/corn3d_group/insseg-pointgroup-v2m1-0-pt3m2-base.py test=True in model dict.

2. run

python tools/test.py --config-file configs/corn3d_group/insseg-pointgroup-v2m1-0-pt3m2-base.py  --options save_path="{weight_path}"  weight="{weight_path}/model_best.pth"

We provide our best model weights here: model_pth

Reference

• Pointcept

Citation

If you find this project useful in your research, please consider cite:

Xin Yang, Teng Miao, Yitong Tao, Bo Zhang, Xiaotong Wu, Xiaodan Han, Jinshi Yu, Yuncheng Zhou, Hanbing Deng, Ying Wang, Tongyu Xu,
PACANet: A Paired-Attention central axis aggregation network for plant population point cloud segmentation and phenotypic trait Extraction—A case study on maize,
Computers and Electronics in Agriculture,
Volume 237, Part B,
2025,
110611,
ISSN 0168-1699,
https://doi.org/10.1016/j.compag.2025.110611.