VMNet: Voxel-Mesh Network for Geodesic-Aware 3D Semantic Segmentation

Created by Zeyu HU

Introduction

This work is based on our paper VMNet: Voxel-Mesh Network for Geodesic-Aware 3D Semantic Segmentation, which appears at the IEEE International Conference on Computer Vision (ICCV) 2021. Update: The TPAMI (ICCV 2021 SI) version has been released.

In recent years, sparse voxel-based methods have become the state-of-the-arts for 3D semantic segmentation of indoor scenes, thanks to the powerful 3D CNNs. Nevertheless, being oblivious to the underlying geometry, voxel-based methods suffer from ambiguous features on spatially close objects and struggle with handling complex and irregular geometries due to the lack of geodesic information. In view of this, we present Voxel-Mesh Network (VMNet), a novel 3D deep architecture that operates on the voxel and mesh representations leveraging both the Euclidean and geodesic information. Intuitively, the Euclidean information extracted from voxels can offer contextual cues representing interactions between nearby objects, while the geodesic information extracted from meshes can help separate objects that are spatially close but have disconnected surfaces. To incorporate such information from the two domains, we design an intra-domain attentive module for effective feature aggregation and an inter-domain attentive module for adaptive feature fusion. Experimental results validate the effectiveness of VMNet: specifically, on the challenging ScanNet dataset for large-scale segmentation of indoor scenes, it outperforms the state-of-the-art SparseConvNet and MinkowskiNet (74.6% vs 72.5% and 73.6% in mIoU) with a simpler network structure (17M vs 30M and 38M parameters).

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{hu2021vmnet,
  title={VMNet: Voxel-Mesh Network for Geodesic-Aware 3D Semantic Segmentation},
  author={Hu, Zeyu and Bai, Xuyang and Shang, Jiaxiang and Zhang, Runze and Dong, Jiayu and Wang, Xin and Sun, Guangyuan and Fu, Hongbo and Tai, Chiew-Lan},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month = {October},
  year = {2021}
}

Installation

• Our code is based on Pytorch. Please make sure CUDA and cuDNN are installed. One configuration has been tested:

  • Python 3.7
  • Pytorch 1.4.0
  • torchvision 0.5.0
  • CUDA 10.0
  • cudatoolkit 10.0.130
  • cuDNN 7.6.5

• VMNet depends on the torch-geometric and torchsparse libraries. Please follow their installation instructions. One configuration has been tested: (Update: higher versions may not work since both torch-geometric and torchsparse have made some changes that are incompatible with the old versions)

  • torch-geometric 1.6.3
  • torchsparse 1.1.0

• We adapted VCGlib (Update: change the link from the official repository to the fork used in DCM-Net, there seems to be a difference in the implementation of tridecimator) to generate pooling trace maps for vertex clustering and quadric error metrics.

  git clone https://github.com/JonasSchult/vcglib
  
  # QUADRIC ERROR METRICS
  cd vcglib/apps/tridecimator/
  qmake
  make
  
  # VERTEX CLUSTERING
  cd ../sample/trimesh_clustering
  qmake
  make
  

  Please add vcglib/apps/tridecimator and vcglib/apps/sample/trimesh_clustering to your environment path variable.

• Other dependencies. One configuration has been tested:

  • open3d 0.9.0
  • plyfile 0.7.3
  • scikit-learn 0.24.0
  • scipy 1.6.0

Data Preparation

• Please refer to https://github.com/ScanNet/ScanNet and https://github.com/niessner/Matterport to get access to the ScanNet and Matterport dataset. Our method relies on the .ply as well as the .labels.ply files. We take ScanNet dataset as example for the following instructions.

• Create directories to store processed data.

  • 'path/to/processed_data/train/'
  • 'path/to/processed_data/val/'
  • 'path/to/processed_data/test/'

• Prepare train data.

  python prepare_data.py --considered_rooms_path dataset/data_split/scannetv2_train.txt --in_path path/to/ScanNet/scans --out_path path/to/processed_data/train/
  

• Prepare val data.

  python prepare_data.py --considered_rooms_path dataset/data_split/scannetv2_val.txt --in_path path/to/ScanNet/scans --out_path path/to/processed_data/val/
  

• Prepare test data.

  python prepare_data.py --test_split --considered_rooms_path dataset/data_split/scannetv2_test.txt --in_path path/to/ScanNet/scans_test --out_path path/to/processed_data/test/
  

Train

• On train/val/test setting.

  CUDA_VISIBLE_DEVICES=0 python run.py --train --exp_name name_you_want --data_path path/to/processed_data
  

• On train+val/test setting (for ScanNet benchmark).

  CUDA_VISIBLE_DEVICES=0 python run.py --train_benchmark --exp_name name_you_want --data_path path/to/processed_data
  

Inference

• Validation. Pretrained model (73.3% mIoU on ScanNet Val). Please download and put into directory check_points/val_split.

  CUDA_VISIBLE_DEVICES=0 python run.py --val --exp_name val_split --data_path path/to/processed_data
  

• Test. Pretrained model (74.6% mIoU on ScanNet Test). Please download and put into directory check_points/test_split. TxT files for benchmark submission will be saved in directory test_results/.

  CUDA_VISIBLE_DEVICES=0 python run.py --test --exp_name test_split --data_path path/to/processed_data
  

Acknowledgements

Our code is built upon torch-geometric, torchsparse and dcm-net.

License

Our code is released under MIT License (see LICENSE file for details).