Code for HongKong-3D datasets Semantic Segmentation

---

            

points with color                             labeled points

---

Requirements

The code is tested on the following environment

• Python 3.9
• Pytorch 1.12.1
• laspy 2.2.2
• open3d 0.15.0
• matplotlib 3.5.1
• numpy 1.24.2
• pathlib 1.0.1
• CUDA 11.3
• cudnn 8.2.1

Dataset

• Original dataset: HongKong-3D(.las/.laz format)
• Processed dataset: HongKong-3D(.ply format)

Usage

Data Preprocessing

• Using /utils/data_prepare_hongkong.py to preprocess the original dataset.
• Using /utils/generate_augmented_pc.py to generate augmented point clouds.

Training

• Using /main_HongKong.py to train the model. You can try like following (more detail in the code):

python main_HongKong.py --name NAME --log_dir LOG_DIR --max_epoch EPOCH --gpu GPU_ID --vla_split VAL_SPLIT

Testing

• Using /test_HongKong.py to test the model. You can try like following (more detail in the code):

python test_HongKong.py --checkpoint_path CHECKPOINT_PATH --name NAME --log_dir LOG_DIR

Pretrained Model

Trained models are provided for use, for details see Pretrained Model

config

• Using /helper_tool.py to set the parameters(including bath_size. etc).

Results

Reference

• Out work are based on the RandLA-Net

Concact Us

Please contact us if you have any questions during use or if you have any other needs.

• Email: [email protected]