Official repository of PanoAVQA: Grounded Audio-Visual Question Answering in 360° Videos (ICCV 2021)

This code is based on following libraries:

• python=3.8
• pytorch=1.7.0 (with cuda 10.2)

To create virtual environment with all necessary libraries:

conda env create -f environment.yml

By default data should be saved under data/feat/{audio,label,visual} directory and logs (w/ cache, checkpoint) are saved under data/{cache,ckpt,log} directory. Using symbolic link is recommended:

ln -s {path_to_your_data_directory} data

We use single TITAN RTX for training, but GPUs with less memory are still doable with smaller batch size (provided precomputed features).

We plan to release the Pano-AVQA dataset public within this year, including Q&A annotation, precomputed features, etc. Please stay tuned!

Default configuration is provided in code/config.py. To run with this configuration:

python cli.py

To run with custom configuration, either modify code/config.py or execute:

python cli.py with {{flags_at_your_disposal}}

Model weight is saved under ./data/log directory. To run inference only:

python cli.py eval with ckpt_file=../data/log/{experiment}/{ckpt}.pth

If you find our work useful in your research, please consider citing:

@InProceedings{Yun2021PanoAVQA,
    author = {Yun, Heeseung and Yu, Youngjae and Yang, Wonsuk and Lee, Kangil and Kim, Gunhee},
    title = {Pano-AVQA: Grounded Audio-Visual Question Answering on 360$^\circ$ Videos},
    booktitle = {ICCV},
    year = {2021}
}

If you have any inquiries, please don't hesitate to contact us via heeseung.yun at vision.snu.ac.kr.