ViewRefer: Grasp the Multi-view Knowledge for 3D Visual Grounding with GPT and Prototype Guidance

Official implementation of 'ViewRefer: Grasp the Multi-view Knowledge for 3D Visual Grounding with GPT and Prototype Guidance'.

The paper has been accepted by ICCV 2023.

News

• We release the GPT-expanded Sr3D dataset and the training code of ViewRefer 📌.

[2023.9] We release AAAI2024 'Point-PEFT', adapting 3D pre-trained Models with 1% parameters to downstream tasks .

[2024.4] We release 'Any2Point', adapting Any-Modality pre-trained Models with 1% parameters to 3D downstream tasks with SOTA performance.

Introduction

ViewRefer is a multi-view framework for 3D visual grounding, which grasps view knowledge to alleviate the challenging view discrepancy issue. For the text and 3D modalities, we respectively introduce LLM-expanded grounding texts and a fusion transformer for capturing multi-view information. We present multi-view prototypes to provide highlevel guidance to our framework, which contributes to superior 3D grounding performance.

Requirements

Please refer to referit3d for the installation and data preparation.

We adopt pre-trained BERT from huggingface. Please install related packages:

pip install transformers

Download the pre-trained BERT, and put them into a folder, noted as PATH_OF_BERT.

Download GPT-expanded Sr3D dataset, and put them into the folder './data'.

Getting Started

Training

• To train on Sr3D dataset, run:

    SR3D_GPT='./referit3d_3dvg/data/Sr3D_release.csv'
    PATH_OF_SCANNET_FILE='./referit3d_3dvg/data/keep_all_points_with_global_scan_alignment.pkl'
    PATH_OF_REFERIT3D_FILE=${SR3D_GPT}
    PATH_OF_BERT='./referit3d_3dvg/data/bert'

    VIEW_NUM=4
    EPOCH=100
    DATA_NAME=SR3D
    EXT=ViewRefer
    DECODER=4
    NAME=${DATA_NAME}_${VIEW_NUM}view_${EPOCH}ep_${EXT}
    TRAIN_FILE=train_referit3d

    python -u ./referit3d_3dvg/scripts/${TRAIN_FILE}.py \
    -scannet-file ${PATH_OF_SCANNET_FILE} \
    -referit3D-file ${PATH_OF_REFERIT3D_FILE} \
    --bert-pretrain-path ${PATH_OF_BERT} \
    --log-dir logs/results/${NAME} \
    --model 'referIt3DNet_transformer' \
    --unit-sphere-norm True \
    --batch-size 24 \
    --n-workers 8 \
    --max-train-epochs ${EPOCH} \
    --encoder-layer-num 3 \
    --decoder-layer-num ${DECODER} \
    --decoder-nhead-num 8 \
    --view_number ${VIEW_NUM} \
    --rotate_number 4 \
    --label-lang-sup True

• Refer to this link for the checkpoint and training log of ViewRefer on Sr3D dataset.

Test

• To test on Sr3D dataset, run:

    SR3D_GPT='./referit3d_3dvg/data/Sr3D_release.csv'
    PATH_OF_SCANNET_FILE='./referit3d_3dvg/data/keep_all_points_with_global_scan_alignment.pkl'
    PATH_OF_REFERIT3D_FILE=${SR3D_GPT}
    PATH_OF_BERT='./referit3d_3dvg/data/bert'

    VIEW_NUM=4
    EPOCH=100
    DATA_NAME=SR3D
    EXT=ViewRefer_test
    DECODER=4
    NAME=${DATA_NAME}_${VIEW_NUM}view_${EPOCH}ep_${EXT}
    TRAIN_FILE=train_referit3d

    python -u ./referit3d_3dvg/scripts/${TRAIN_FILE}.py \
    --mode evaluate \
    -scannet-file ${PATH_OF_SCANNET_FILE} \
    -referit3D-file ${PATH_OF_REFERIT3D_FILE} \
    --bert-pretrain-path ${PATH_OF_BERT} \
    --log-dir logs/results/${NAME} \
    --resume-path "./checkpoints/best_model.pth"\
    --model 'referIt3DNet_transformer' \
    --unit-sphere-norm True \
    --batch-size 24 \
    --n-workers 8 \
    --max-train-epochs ${EPOCH} \
    --encoder-layer-num 3 \
    --decoder-layer-num ${DECODER} \
    --decoder-nhead-num 8 \
    --view_number ${VIEW_NUM} \
    --rotate_number 4 \
    --label-lang-sup True

Acknowledgement

This repo benefits from ReferIt3D and MVT-3DVG. Thanks for their wonderful works.

Citation

@article{guo2023viewrefer,
  title={ViewRefer: Grasp the Multi-view Knowledge for 3D Visual Grounding with GPT and Prototype Guidance},
  author={Guo, Ziyu and Tang, Yiwen and Zhang, Renrui and Wang, Dong and Wang, Zhigang and Zhao, Bin and Li, Xuelong},
  journal={arXiv preprint arXiv:2303.16894},
  year={2023}
}

Contact

If you have any questions about this project, please feel free to contact [email protected].