HCNQA: Enhancing 3D VQA with Hierarchical Concentration Narrowing Supervision

TL;DR HCNQA improves 3D Visual Question-Answering by hierarchically supervising reasoning phases (BoI, OoI, OoT) to suppress shortcuts, achieving better accuracy and robustness.

Installation

conda create -n hcnqa python=3.9 -y
conda activate hcnqa
pip install -r requirements.txt
cd model/vision/pointnet2
pip install .

Data Preparation

1. Please refer to the tutorial of 3D-VisTA, download and extract the ScanQA dataset.
2. download the annotation of our work from coarse_ground_train_5.json and coarse_ground_val_5.json.

Checkpoints

1. Download the checkpoint for the language encoder (bert-base-uncased).
2. Download the checkpoint for our model from eqa_235_5x5_ft5_2389.pth.

Training

python run.py --config project/vista/scanqa_train.yml

Evaluation

python run.py --config project/vista/scanqa_eval.yml

License & Acknowledgements

This repository is released under the MIT License (see LICENSE).

The whole codebase inherits heavily from the open-sourced 3D-VisTA project.

We gratefully acknowledge their clean design and utilities.

Citation

If you find this project useful in your research, please consider citing:

@misc{zhou2025hcnqaenhancing3dvqa,
      title={HCNQA: Enhancing 3D VQA with Hierarchical Concentration Narrowing Supervision}, 
      author={Shengli Zhou and Jianuo Zhu and Qilin Huang and Fangjing Wang and Yanfu Zhang and Feng Zheng},
      year={2025},
      eprint={2507.01800},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.01800}, 
}