This is the code for the paper "Learning Point-Language Hierarchical Alignment for 3D Visual Grounding".

Our method is the 1st method on the ScanRefer benchmark (2022/10 - 2023/3) and is the winner of the ECCV 2022 2nd Workshop on Language for 3D Scenes.

1. Download the ScanRefer dataset and unzip it under data/.
2. Downloadand the preprocessed GLoVE embeddings and put them under data/.
3. Download the ScanNetV2 dataset and put scans/ under data/scannet/scans/.
4. Pre-process ScanNet data.

   cd data/scannet
   python batch_load_scannet_data.py
   

pip install -r requirements.txt
cd lib/pointnet2
python setup.py install

Set the correct project path in lib/config.py.

Using --tag to name your experiment, and the training snapshots and results will be put in outputs/TAG_NAME_[timestamp]

CUDA_VISIABLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch \
           --master_port 19999 --nproc_per_node 8 ./scripts/train_dist.py \
           --fuse_with_key  --use_spa  --sent_aug  \
           --use_color --use_normal \
           --tag TAG_NAME

CUDA_VISIABLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch 
           --master_port 19998 --nproc_per_node 1 ./benchmark/predict.py  
           --fuse_with_key  --use_spa --use_color --use_normal 
           --no_nms --pred_split val --folder TAG_NAME

We thank a lot for the codebases of ScanRefer, 3DVG-Transformer, GroupFree.