SCA3D: Enhancing Cross-modal 3D Retrieval via 3D Shape and Caption Paired Data Augmentation

Junlong Ren* · Hao Wu* · Hui Xiong · Hao Wang✉

The Hong Kong University of Science and Technology (GuangZhou)

(* Equal Contribution)

ICRA 2025

Preparation

Downloading Datasets

The point clouds, text captions, and segmentation annotations are provided by Parts2Words. You can download the files here.

Checkpoints

We also provide the pre-trained model weights in Google Drive.

Quick Start

# Train
CUDA_VISIBLE_DEVICES=0 python train.py --config config/SCA3D.yaml

# Eval
CUDA_VISIBLE_DEVICES=0 python val.py --config config/SCA3D.yaml

Citation

If you feel this project helpful to your research, please cite our work.

@article{ren2025sca3d,
  title={SCA3D: Enhancing Cross-modal 3D Retrieval via 3D Shape and Caption Paired Data Augmentation},
  author={Ren, Junlong and Wu, Hao and Xiong, Hui and Wang, Hao},
  journal={arXiv preprint arXiv:2502.19128},
  year={2025}
}

Acknowledgement

• This work is built on Parts2Words and we borrow some codes from CoVR-ECDE. Thanks for these great works.