Learning Gaussian Instance Segmentation in Point Clouds

Introduction

Shih-Hung Liu, Shang Yi Yu, Shao-Chi Wu, Hwann-Tzong Chen, Tyng-Luh Liu

(1) Setup

ubuntu 16.04 + cuda 10.1

python 3.6

pytorch 1.5.1

scipy 1.3

h5py 2.9

open3d-python 0.3.0

(2) Data

S3DIS: we use the same data released by JSIS3D. You can download the data into the ./data_s3dis

ScnaNet: you can download the ScanNet data in ScanNet.

(3) Train/test

python train.py

python main_eval.py

(4) Compilation

1. Compiling the pointnet++ module

cd Pointnet2.PyTorch/pointnet2

python setup.py install

2. You also need to compiling SCN for semantic prediction

The environment is based on facebookresearch/SparseConvNet

(5) Quantitative Results on ScanNet

(6) Pre-trained model

The pretrained GICN on S3dis dataset is in ./experiment

Evaluation on Area5:

-precision : 0.6348

-recall : 0.4669

(7) Acknowledgements

Pointnet++ is based on sshaoshuai/Pointnet2.PyTorch