PointNet Segmentation on S3DIS Dataset

Requirements

• Pytorch with your favourite cuda
• Open3D

Instructions

Dataset

Dataset Download the dataset 'Stanford3dDataset_v1.2_Aligned_Version' in 'point_net' folder. Also for testing on new scan (much different than dataset), I suggest to download scans from here, particularly scans in here, and place them in 'point_net' folder

Main Training

cd point_net
# convert the S3DIS dataset into hd5 format
python s3dis_reducer.py
# train the pointnet on training and validation data of S3DIS
python main_train.py
# test pointnet on testing data of S3DIS
python main_test.py
cd ..

Plotting

cd point_net

# plot train and validation metrics and save them as png
python plot_train_val_metrics.py

# plot predictions and ground truth pointclouds of testing S3DIS dataset and save them as png
python pointclouds_2_png.py

# compare the ground truth and predictions of testing S3DIS dataset over epochs, in single png
python compare_png.py
cd ..

Testing on Unseen Data

cd point_net
python test_new.py --path <path to .ptx file>

Ground Truth vs Predicted Point Cloud

Ground Truth


Predicted

Segmentation Performance Over Epochs

Testing on Unseen Point Cloud