Awesome

TransRefer3D: Entity-and-Relation Aware Transformer for Fine-Grained 3D Visual Grounding

TransRefer3D: Entity-and-Relation Aware Transformer for Fine-Grained 3D Visual Grounding [ACM MM'21] [ArXiv]

Dailan He*, Yusheng Zhao*, Junyu Luo*, Tianrui Hui, Shaofei Huang, Aixi Zhang and Si Liu†.

🔥 Checkout our latest works in 3D Visual Grounding at 3D-SPS.🔥

Introduction

In this work, we exploit Transformer for its natural suitability on permutation-invariant 3D point clouds data and propose a TransRefer3D network to extract entity-and-relation aware multimodal context among objects for more discriminative feature learning. Concretely, we devise an Entity-aware Attention (EA) module and a Relation-aware Attention (RA) module to conduct fine-grained cross-modal feature matching.

Installation and Data Preparation

Please refer the installation and data preparation from referit3d.

Training

To train on either Nr3d or Sr3d dataset, use the following commands

    python3 train_referit3d.py \
        --gpu 0 \
        --depth 4 \
        --experiment-tag $EXERPIMENT_TAG$ \
        --model transrefer3d \
        -scannet-file $PATH_OF_SCANNET_FILE$ \
        -referit3D-file $PATH_OF_REFERIT3D_FILE$ \
        --log-dir $LOG_DIR$ \
        --n-workers 4

feel free to change the number of workers to match your #CPUs and RAM size.

To train nr3d in joint with sr3d, add the following argument

    --augment-with-sr3d sr3d_dataset_file.csv

Evaluation

To evaluate on either Nr3d or Sr3d dataset, use the following commands

python3 train_referit3d.py \
    --mode evaluate \
    --model transrefer3d \
    -scannet-file $PATH_OF_SCANNET_FILE$ \
    -referit3d-file $PATH_OF_REFERIT3D_FILE$ \
    --resume-path $PATH_OF_MODEL$ \
    --n-workers 4

To evaluate on joint trained model, add the following argument to the above command

    --augment-with-sr3d sr3d_dataset_file.csv

Pretrained Models

On Nr3D: [Google Drive] [Baidu Pan]

On Sr3D: [Google Drive] [Baidu Pan]

Architecture of the proposed TransRefer3D. The blue arrows indicates the flow of visual features, while the red arrows showcases the flow of linguistic features.

Citation

Please consider citing the corresponding publication if you find this work or code is helpful in your research,

@inproceedings{transrefer3d,
    title={TransRefer3D: Entity-and-Relation Aware Transformer for Fine-Grained 3D Visual Grounding},
    author={He, Dailan and Zhao, Yusheng and Luo, Junyu and Hui, Tianrui and Huang, Shaofei and Zhang, Aixi and Liu, Si},
    booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
    year={2021}
}

Credits

The project is built based on the following repository:

ReferIt3D