NEAT: Distilling 3D Wireframes from Neural Attraction Fields (To be updated)

Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, Yujun Shen

2024

Preprint / Code / Video / Processed Data (4.73 GB) / Precomputed Results (3.01 GB)

git clone https://github.com/cherubicXN/neat.git --recursive

conda create -n neat python=3.10
conda activate neat

pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117

cd third-party/hawp
pip install -e .
cd ../..

pip install -r requirements.txt

• Step 1: Training or Optimization

  python training/exp_runner.py --conf confs/abc-neat-a.conf --nepoch 2000 --tbvis # --tbvis will use tensorboard for visualization
  

• Step 2: Finalize the NEAT wireframe model

  python neat-final-parsing.py --conf ../exps/abc-neat-a/{timestamp}/runconf.conf --checkpoint 1000
  

  After running the above command line, you will get 4 files at ../exps/abc-neat-a/{timestamp}/wireframes with the prefix of {epoch}-{hash}*, where {epoch} is the checkpoint you evaluated and {hash} is an hash of hyperparameters for finalization.

  The four files are with the different suffix strings:

  • {epoch}-{hash}-all.npz stores the all line segments from the NEAT field,
  • {epoch}-{hash}-wfi.npz stores the initial wireframe model without visibility checking, containing some artifacts in terms of the wireframe edges,
  • {epoch}-{hash}-wfi_checked.npz stores the wireframe model after visibility checking to reduce the edge artifacts,
  • {epoch}-{hash}-neat.pth stores the above three files and some other information in the pth format.

• Step 3: Visualize the 3D wireframe model by

  python visualization/show.py --data ../exps/abc-neat-a/{timestamp}/wireframe/{filename}.npz 
  

  

  • Currently, the visualization script only supports the local run.
  • The open3d (v0.17) plugin for tensorboard is slow

• Precomputed results can be downloaded from url-results
• Processed data can be downloaded from url-data, which are organized with the following structure:

data
├── BlendedMVS
│   ├── process.py
│   ├── scan11
│   ├── scan13
│   ├── scan14
│   ├── scan15
│   └── scan9
├── DTU
│   ├── bbs.npz
│   ├── scan105
│   ├── scan16
│   ├── scan17
│   ├── scan18
│   ├── scan19
│   ├── scan21
│   ├── scan22
│   ├── scan23
│   ├── scan24
│   ├── scan37
│   ├── scan40
│   └── scan65
├── abc
│   ├── 00004981
│   ├── 00013166
│   ├── 00017078
│   └── 00019674
└── preprocess
    ├── blender.py
    ├── extract_monocular_cues.py
    ├── monodepth.py
    ├── normalize.py
    ├── normalize_cameras.py
    ├── parse_cameras_blendedmvs.py
    └── readme.md

• Evaluation code (To be updated)

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

@article{NEAT-arxiv,
  author       = {Nan Xue and
                  Bin Tan and
                  Yuxi Xiao and
                  Liang Dong and
                  Gui{-}Song Xia and
                  Tianfu Wu and
                  Yujun Shen
                 },
  title        = {Volumetric Wireframe Parsing from Neural Attraction Fields},
  journal      = {CoRR},
  volume       = {abs/2307.10206},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2307.10206},
  doi          = {10.48550/arXiv.2307.10206},
  eprinttype    = {arXiv},
  eprint       = {2307.10206}
}

This project is built on volsdf. We also thank the four anonymous reviewers for their feedback on the paper writing, listed as follows (copied from the CMT system):