This project is based on our CVPR 2020 paper,Fusion-Aware Point Convolution for Online Semantic 3D Scene Segmentation
We propose a novel fusionaware 3D point convolution which operates directly on the geometric surface being reconstructed and exploits effectively the inter-frame correlation for high quality 3D feature learning.
This code is based on PyTorch and needs open3D for convenient visualization
Our code has been tested with Python 3.7.6, PyTorch 1.1.0, open3d 0.9.0, CUDA 10.0 on Ubuntu 16.04.
We use the ScanNetv2 as our test dataset. If you want to test all the data, you can download the ScanNetV2 dataset from here. For a quick visulazation test, we provide several pre-proessing scenes of the test set sequence. Put the scene.h5 in path/data.
We also provide the pre-trained weights for ScanNet benchmark, you can download from here. After finishing the download, put the weights in path/weight.
We have already intergrate the open3d for visulizaiton, you can run the command below:
python vis_sequence.py --weight2d_path=weight_path/weight2d_name --weight3d_path=weight_path/weight3d_name --gpu=0 --use_vis=1 --scene_path=scene_path/scene_name
The complete segmentation result will be generated in result.ply.
We achieve the a test demo for global-local tree visulizaiton only. Run the command below to see the processing of the tree built.
python vis_sequence.py --use_vis=1 --scene_path=scene_path/scene_name
The complete result will be generated in result_GLtree.ply.
If you find our work useful in your research, please consider citing:
@article{zhang2020fusion,
title={Fusion-Aware Point Convolution for Online Semantic 3D Scene Segmentation},
author={Zhang, Jiazhao and Zhu, Chenyang and Zheng, Lintao and Xu, Kai},
journal={arXiv preprint arXiv:2003.06233},
year={2020}
}
Code is inspired by Red-Black-Tree and FuseNet_PyTorch.
If you have any questions, please email Jiazhao Zhang at zhngjizh@gmail.com.