Awesome

Disentangled Instance Mesh Reconstruction (ECCV 2022)

This repository contains the official implementation for the paper: Point Scene Understanding via Disentangled Instance Mesh Reconstruction.

Project Page | Arxiv | Data

Installation

Clone the repository:

git clone --recursive  https://github.com/ashawkey/dimr
cd dimr

pip install -r requirements.txt

Install dependent libraries:

The repository depends on a modified spconv from pointgroup for sparse convolution, which requires CUDA version < 11 and pytorch < 1.5.

• spconv

  cd lib/spconv
  python setup.by bdist_wheel
  cd dist
  # may need to change the filename
  pip install spconv-1.0-cp37-cp37m-linux_x86_64.whl
  

• pointgroup_ops

  cd lib/pointgroup_ops
  python setup.py develop
  

• bspt_ops

  cd lib/bspt
  python setup.py develop
  

• light-field-distance

  cd lib/light-field-distance
  python setup.py develop
  

Data Preparation

Full folder structure

.
├──datasets
│   ├── scannet
│   │   ├── scans # scannet scans
│   │   │   ├── scene0000_00 # only these 4 files are used.
│   │   │   │   ├── scene0000_00.txt
│   │   │   │   ├── scene0000_00_vh_clean_2.ply
│   │   │   │   ├── scene0000_00.aggregation.json
│   │   │   │   ├── scene0000_00_vh_clean_2.0.010000.segs.json
│   │   │   ├── ......
│   │   │   ├── scene0706_00
│   │   ├── scan2cad # scan2cad, only the following 1 file is used.
│   │   │   ├── full_annotations.json
│   │   ├── scannetv2-labels-combined.tsv # scannet label mappings
│   │   ├── processed_data # preprocessed data
│   │   │   ├── scene0000_00 
│   │   │   │   ├── bbox.pkl
│   │   │   │   ├── data.npz
│   │   │   ├── ......
│   │   │   ├── scene0706_00
│   │   ├── rfs_label_map.csv # generated label mappings
│   ├── ShapeNetCore.v2 # shapenet core v2 dataset
│   │   ├── 02954340
│   │   ├── ......
│   │   ├── 04554684
│   ├── ShapeNetv2_data # preprocessed shapenet dataset
│   │   ├── watertight_scaled_simplified
│   ├── bsp # the pretrained bsp model
│   │   ├── zs
│   │   ├── database_scannet.npz
│   │   ├── model.pth
│   ├── splits # data splits
│   │   ├── train.txt
│   │   ├── val.txt
│   │   ├── test.txt

Prepare the data

• download the preprocesssed data here (~3.3G) and label map here, and put them under ./datasets/scannet.

  If you want to process the data by yourself, please:

  • download the ScanNet dataset, Scan2CAD dataset, and the ShapeNet dataset, and put them to the corresponding locations.

  • preprocess ScanNet data by:

    python data/generate_data_relabel.py
    

    By default it launches 16 processes to accelerate processing, which will finish in about 10 minutes. Due to mismatching of instance labels and CAD annotations, it may log some warnings, but will not affect the process. It will generate the label map ./datasets/scannet/rfs_label_map.csv, and save data.npz, bbox.pkl for each scene under ./datasets/scannet/processed_data/.

• download the preprocessed ShapeNet (simplified watertight mesh) following RfDNet into ShapeNetv2_data, only the watertight_scaled_simplified is used for the mesh retrieval and evaluation.

• download the pretrained BSP-Net checkpoint and extracted GT latent codes here.

  If you want to generate them by yourself, please check the BSP_CVAE repository to generate the ground truth latent shape codes (zs folder), the pretrained model (model.pth), and the assistant code database (database_scannet.npz).

Training

# train phase 1 (point-wise)
python train.py --config config/rfs_phase1_scannet.yaml

# train phase 2 (proposal-wise)
python train.py --config config/rfs_phase2_scannet.yaml

Please check the config files for more options.

Testing

Generate completed instance meshes:

# test after training phase 2
python test.py --config config/rfs_phase2_scannet.yaml
# example path for the meshes: ./exp/scannetv2/rfs/rfs_phase2_scannet/result/epoch256_nmst0.3_scoret0.05_npointt100/val/trimeshes/

# test with a speficied checkpoint
python test.py --config config/rfs_pretrained_scannet.yaml --pretrain ./checkpoint.pth

We provide the pretrained model here.

To visualize the intermediate point-wise results:

python util/visualize.py --task semantic_gt --room_name all
python util/visualize.py --task instance_gt --room_name all
# after running test.py, may need to change `--result_root` to the output directory, check the script for more details.
python util/visualize.py --task semantic_pred --room_name all
python util/visualize.py --task instance_pred --room_name all

Evaluation

We provide 4 metrics for evaulation the instance mesh reconstruction quality. For the IoU evaluation, we rely on binvox to voxelize meshes (via trimesh's API), so make sure it can be found in the system path.

## first, prepare GT instance meshes 
python data/scannetv2_inst.py # prepare at "./datasets/gt_meshes"

## assume the generated meshes are under "./pred_meshes"
# IoU
python evaluation/iou/eval.py ./datasets/gt_meshes ./pred_meshes

# CD
python evaluation/cd/eval.py ./datasets/gt_meshes ./pred_meshes

# LFD
python evaluation/lfd/eval.py ./datasets/gt_meshes ./pred_meshes

# PCR
python evaluation/pcr/eval.py ./pred_meshes

We provide the meshes used in our evaluation for reproduction here, it includes the output meshes from RfD-Net, Ours, Ours-projection, and Ours-retrieval.

The GT meshes can also be found here.

Citation

If you find our work useful, please use the following BibTeX entry:

@article{tang2022point,
  title={Point Scene Understanding via Disentangled Instance Mesh Reconstruction},
  author={Tang, Jiaxiang and Chen, Xiaokang and Wang, Jingbo and Zeng, Gang},
  journal={arXiv preprint arXiv:2203.16832},
  year={2022}
}

Acknowledgement

We would like to thank RfD-Net and pointgroup authors for open-sourcing their great work!