This repository is for the HAIS_2GNN research project.

Tao Gu, Yue Chen

The motivation of this project is to improve the accuracy of 3D visual grounding. In this report, we propose a new model, named HAIS_2GNN based on the InstanceRefer model, to tackle the problem of insufficient connections between instance proposals. Our model incorporates a powerful instance segmentation model HAIS and strengthens the instance features by the structure of graph and attention, so that the text and point cloud can be better matched together. Experiments confirm that our method outperforms the InstanceRefer on ScanRefer validation datasets. Link to the technical report

The code is tested on Ubuntu 20.04.3 LTS with Python 3.9.7 PyTorch 1.10.1 CUDA 11.3.1 installed.

conda install pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.3 -c pytorch

Install the necessary packages listed out in requirements.txt:

pip install -r requirements.txt

After all packages are properly installed, please run the following commands to compile the torchsaprse v1.4.0:

sudo apt-get install libsparsehash-dev
pip install --upgrade git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0

Before moving on to the next step, please don't forget to set the project root path to the CONF.PATH.BASE in lib/config.py.

1. Download the ScanRefer dataset and unzip it under data/.
2. Downloadand the preprocessed GLoVE embeddings (~990MB) and put them under data/.
3. Download the ScanNetV2 dataset and put (or link) scans/ under (or to) data/scannet/scans/ (Please follow the ScanNet Instructions for downloading the ScanNet dataset). After this step, there should be folders containing the ScanNet scene data under the data/scannet/scans/ with names like scene0000_00
4. Used official and pre-trained HAIS generate panoptic segmentation in PointGroupInst/. We will provide the pre-trained data soon.
5. Pre-processed instance labels, and new data should be generated in data/scannet/pointgroup_data/

cd data/scannet/
python prepare_data.py --split train --pointgroupinst_path [YOUR_PATH]
python prepare_data.py --split val   --pointgroupinst_path [YOUR_PATH]
python prepare_data.py --split test  --pointgroupinst_path [YOUR_PATH]

Finally, the dataset folder should be organized as follows.

InstanceRefer
├── data
│   ├── glove.p
│   ├── ScanRefer_filtered.json
│   ├── ...
│   ├── scannet
│   │  ├── meta_data
│   │  ├── pointgroup_data
│   │  │  ├── scene0000_00_aligned_bbox.npy
│   │  │  ├── scene0000_00_aligned_vert.npy
│   │  ├──├──  ... ...

Train the InstanceRefer model. You can change hyper-parameters in config/InstanceRefer.yaml:

python scripts/train.py --log_dir HAIS_2GNN

You need specific the use_checkpoint with the folder that contains model.pth in config/InstanceRefer.yaml and run with:

python scripts/eval.py

• Add pre-trained HAIS dataset.
• Release pre-trained model.
• Merge HAIS in an end-to-end manner.
• Upload to ScanRefer benchmark

02/09/2022: Released HAIS_2GNN

This work is a research project conducted by Tao Gu and Yue Chen for ADL4CV:Visual Computing course at the Technical University of Munich.

We acknowledge that our work is based on ScanRefer, InstanceRefer, HAIS, torchsaprse, and pytorch_geometric.

This repository is released under MIT License (see LICENSE file for details).