Transfer pointnet2 to binary classification for plane and non-plane.

Requirements

• Python 3.7
• Pytorch 1.13.1
• os
• sys
• argparse
• logging
• importlib
• datetime
• shutil
• pathlib
• tqdm
• numpy

Installation

git clone https://github.com/xiaotaiyangzmh/PointNet2_plane.git
cd PointNet2_plane/

Training

If the training uses transfer learning Run the following command to train a plane segmentation model with real scene

python train_semseg.py --model pointnet2_sem_seg --train_path "./data_scene/crop_data" --test_path "./data_scene/crop_testdata" --batch_size 128 --epoch 64 --log_dir pointnet2_real_data --transfer

If the model is trained from scratch, do not add argument --transfer, for example

python train_semseg.py --model pointnet2_sem_seg --train_path "./data_scene/crop_data" --test_path "./data_scene/crop_testdata" --batch_size 128 --epoch 64 --log_dir pointnet2_real_data

Testing

Run the following command to test the model with the labelled whole scene

python test_labelled.py --log_dir pointnet2_real_data --test_path "./data_scene/crop_testdata" --visual --model_epoch 63

Run the following command to test the model with the unlabelled whole scene

python draw_results.py --log_dir pointnet2_real_data_0823_xyz_w4 --target_dir crop_testdata --prediction_dir eval_labelled_63

Testing

Run the following command to plot figures ''' python draw_results.py --exp_dir plane_seg --log_dir pointnet2_real_data '''

Exporting to c++

python export_pytorch_jit.py --log_dir pointnet2_synthetic_data python export_pytorch_jit.py --log_dir pointnet2_synthetic_combined_data

Reference

charlesq34/PointNet++
yanx27/Pointnet_Pointnet2_pytorch