EDA: Explicit Text-Decoupling and Dense Alignment for 3D Visual Grounding (CVPR2023)

By Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, Jian Zhang*
This repo is the official implementation of "EDA: Explicit Text-Decoupling and Dense Alignment for 3D Visual Grounding". CVPR2023 | arXiv | Code

0. Installation

• (1) Install environment with environment.yml file:

  conda env create -f environment.yml --name EDA
  

  • or you can install manually:

    conda create -n EDA python=3.7
    conda activate EDA
    conda install pytorch==1.9.0 torchvision==0.10.0 cudatoolkit=11.1 -c pytorch -c nvidia
    pip install numpy ipython psutil traitlets transformers termcolor ipdb scipy tensorboardX h5py wandb plyfile tabulate
    

• (2) Install spacy for text parsing

  pip install spacy
  # 3.3.0
  pip install https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.3.0/en_core_web_sm-3.3.0.tar.gz
  

• (3) Compile pointnet++

  cd ~/EDA
  sh init.sh
  

1. [TODO] Quick visualization demo

• [ ] Visualization
• [ ] Text-decoupling demo

2. Data preparation

The final required files are as follows:

├── [DATA_ROOT]
│	├── [1] train_v3scans.pkl # Packaged ScanNet training set
│	├── [2] val_v3scans.pkl   # Packaged ScanNet validation set
│	├── [3] ScanRefer/        # ScanRefer utterance data
│	│	│	├── ScanRefer_filtered_train.json
│	│	│	├── ScanRefer_filtered_val.json
│	│	│	└── ...
│	├── [4] ReferIt3D/        # NR3D/SR3D utterance data
│	│	│	├── nr3d.csv
│	│	│	├── sr3d.csv
│	│	│	└── ...
│	├── [5] group_free_pred_bboxes/  # detected boxes (optional)
│	├── [6] gf_detector_l6o256.pth   # pointnet++ checkpoint (optional)
│	├── [7] roberta-base/     # roberta pretrained language model
│	├── [8] checkpoints/      # EDA pretrained models

• [1] [2] Prepare ScanNet Point Clouds Data
  • 1) Download ScanNet v2 data. Follow the ScanNet instructions to apply for dataset permission, and you will get the official download script download-scannet.py. Then use the following command to download the necessary files:

    python2 download-scannet.py -o [SCANNET_PATH] --type _vh_clean_2.ply
    python2 download-scannet.py -o [SCANNET_PATH] --type _vh_clean_2.labels.ply
    python2 download-scannet.py -o [SCANNET_PATH] --type .aggregation.json
    python2 download-scannet.py -o [SCANNET_PATH] --type _vh_clean_2.0.010000.segs.json
    python2 download-scannet.py -o [SCANNET_PATH] --type .txt
    

    where [SCANNET_PATH] is the output folder. The scannet dataset structure should look like below:

    ├── [SCANNET_PATH]
    │   ├── scans
    │   │   ├── scene0000_00
    │   │   │   ├── scene0000_00.txt
    │   │   │   ├── scene0000_00.aggregation.json
    │   │   │   ├── scene0000_00_vh_clean_2.ply
    │   │   │   ├── scene0000_00_vh_clean_2.labels.ply
    │   │   │   ├── scene0000_00_vh_clean_2.0.010000.segs.json
    │   │   ├── scene.......
    

  • 2) Package the above files into two .pkl files(train_v3scans.pkl and val_v3scans.pkl):

    python Pack_scan_files.py --scannet_data [SCANNET_PATH] --data_root [DATA_ROOT]
    

• [3] ScanRefer: Download ScanRefer annotations following the instructions HERE. Unzip inside [DATA_ROOT].
• [4] ReferIt3D: Download ReferIt3D annotations following the instructions HERE. Unzip inside [DATA_ROOT].
• [5] group_free_pred_bboxes: Download object detector's outputs. Unzip inside [DATA_ROOT]. (not used in single-stage method)
• [6] gf_detector_l6o256.pth: Download PointNet++ checkpoint into [DATA_ROOT].
• [7] roberta-base: Download the roberta pytorch model:

  cd [DATA_ROOT]
  git clone https://huggingface.co/roberta-base
  cd roberta-base
  rm -rf pytorch_model.bin
  wget https://huggingface.co/roberta-base/resolve/main/pytorch_model.bin
  

• [8] checkpoints: Our pre-trained models (see next step).

3. Models

*: This model is also used to evaluate the new task of grounding without object names, with performances of 26.5% and 21.6% for [email protected] and [email protected].
1: The log of the performance we reported in the paper.
2: The log of the performance we retrain the model with this open-released repository.
Note: To find the overall performance, please refer to issue3.

4. Training

• Please specify the paths of --data_root, --log_dir, --pp_checkpoint in the train_*.sh script first. We use four or two 24-GB 3090 GPUs for training with a batch size of 12 by default.
• For ScanRefer training

  sh scripts/train_scanrefer.sh
  

• For ScanRefer (single stage) training

  sh scripts/train_scanrefer_single.sh
  

• For SR3D training

  sh scripts/train_sr3d.sh
  

• For NR3D training

  sh scripts/train_nr3d.sh
  

5. Evaluation

• Please specify the paths of --data_root, --log_dir, --checkpoint_path in the test_*.sh script first.
• For ScanRefer evaluation

  sh scripts/test_scanrefer.sh
  

  • New task: grounding without object names. Please first download our new annotation, then give the path of --wo_obj_name in the script and run:

    sh scripts/test_scanrefer_wo_obj_name.sh
    

• For ScanRefer (single stage) evaluation

  sh scripts/test_scanrefer_single.sh
  

• For SR3D evaluation

  sh scripts/test_sr3d.sh
  

• For NR3D evaluation

  sh scripts/test_nr3d.sh
  

6. Acknowledgements

We are quite grateful for BUTD-DETR, GroupFree, ScanRefer, and SceneGraphParser.

7. Citation

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

@inproceedings{wu2022eda,
  title={EDA: Explicit Text-Decoupling and Dense Alignment for 3D Visual Grounding},
  author={Wu, Yanmin and Cheng, Xinhua and Zhang, Renrui and Cheng, Zesen and Zhang, Jian},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2023}
}

8. Contact

If you have any question about this project, please feel free to contact Yanmin Wu: wuyanminmax[AT]gmail.com