Created by Na Zhao from National University of Singapore

This repository contains the PyTorch implementation for our ICPR 2020 Paper "PS²Net: A Locally and Globally Aware Network for Point-Based SemanticSegmentation" by Na Zhao, Tat Seng Chua, Gim Hee Lee [arXiv]

In this paper, we present the PS²-Net - a locally and globally aware deep learning framework for semantic segmentation on 3D scene-level point clouds. In order to deeply incorporate local structures and global context to support 3D scene segmentation, our network is built on four repeatedly stacked encoders, where each encoder has twobasic components: EdgeConv that captures local structures and NetVLAD that models global context. Different from existing state-of-the-art methods for point-based scene semantic segmentation that either violate or do not achieve permutation invariance, our PS²-Net is designed to be permutation invariant which is an essential property of any deep network used to process unordered point clouds. We further provide theoretical proof to guarantee the permutation invariance property of our network. We perform extensive experiments on two large-scale 3D indoor scene datasets and demonstrate that our PS²-Net is able to achieve state-of-the-art performances as compared to existing approaches.

• Install python --This repo is tested with python 3.6.5.
• Install pytorch with CUDA -- This repo is tested with torch 0.4.0, CUDA 9.0. It may wrk with newer versions, but that is not gauranteed.
• Install faiss with CPU version by conda install faiss-cpu -c pytorch -- This repo is tested with faiss 1.4.0
• Install dependencies

  pip install -r requirements.txt
  

For S3DIS, follow the README under ./preprocess/s3dis folder.

For ScanNet, follow the README under ./preprocess/scannet folder.

We use visdom for visualization. Loss values and performance are plotted in real-time. Please start the visdom server before training: python -m visdom.server

The visualization results can be viewed in browser with the address of: http://localhost:8097.

• train on each area:

  python main_P1/train.py --dataset_name S3DIS --data_dir ./datasets/S3DIS/P1/ --classes 13 --input_feat 9 --log_dir $LOG_DIR  --test_area $Area_Index
  

• test on the corresponding area:

  python main_P1/test.py --dataset_name S3DIS --data_dir ./datasets/S3DIS/P1/ --classes 13 --input_feat 9 --log_dir $LOG_DIR  --checkpoint $CHECKPOINT_FILENAME --test_area $Area_Index
  

• train on each area:

  python main_P2/train.py --dataset_name S3DIS --dataset_size 114004 --data_dir ./datasets/S3DIS/P2/ --classes 13 --input_feat 6 --log_dir $LOG_DIR  --test_area $Area_Index
  

• test on the corresponding area:

  python main_P2/inference.py --dataset_name S3DIS --data_dir ./datasets/S3DIS/P2/ --classes 13 --input_feat 6 --log_dir $LOG_DIR  --checkpoint $CHECKPOINT_FILENAME --test_area $Area_Index
  python main_P2/eval_s3dis.py --datafolder ./datasets/S3DIS/P2/ --test_area $Area_Index
  

Note that these commands are for training and evaluating only one area (specified by --test_area $Area_Index option) validation. Please iterate --test_area option to obtain results on other areas. The final result is computed based on 6-fold cross validation.

• train:

  python main_P1/train.py --dataset_name ScanNet --data_dir ./datasets/ScanNet/P3/ --classes 21 --input_feat 3 --log_dir $LOG_DIR 
  

• test:

  python main_P1/test.py --dataset_name ScanNet --data_dir ./datasets/ScanNet/P3/ --classes 21 --input_feat 3 --log_dir $LOG_DIR  --checkpoint $CHECKPOINT_FILENAME
  

• train:

  python main_P2/train.py --dataset_name ScanNet --dataset_size 93402 --data_dir ./datasets/ScanNet/P2/ --classes 21 --input_feat 3 --log_dir $LOG_DIR  
  

• test:

  python main_P2/inference.py --dataset_name ScanNet --data_dir ./datasets/ScanNet/P2/ --classes 21 --input_feat 3 --log_dir $LOG_DIR  --checkpoint $CHECKPOINT_FILENAME 
  python main_P2/eval_scannet.py --datafolder ./datasets/ScanNet/P2/test --picklefile ./datasets/ScanNet/P3/scannet_test.pickle
  

Please cite our paper if it is helpful to your research:

@inproceedings{zhao2020ps,
title={PS\^{}2-Net: A Locally and Globally Aware Network for Point-Based Semantic Segmentation},
author={Zhao, Na and Chua, Tat-Seng and Lee, Gim Hee},
booktitle={Proceedings of the 25th International Conference on Pattern Recognition (ICPR)},
pages={723--730},
year={2020}
}

Our implementation leverages on the source code or data from the following repositories:

• PointNet
• PointNet++
• PointCNN
• SO-Net