AVQACL: A Novel Benchmark for Audio-Visual Question Answering Continual Learning（CVPR2025）

In this paper, a novel benchmark for audio-visual question answering continual learning (AVQACL) is introduced, aiming to study fine-grained scene understanding and spatial-temporal reasoning in videos under a continual learning setting.

Environment

Python 3.12.4

Pytorch 2.4.0

To setup the environment, please run

pip install -r requirements.txt

Datasets

Split-AVQA and Split-MUSIC-AVQA

The feature data extracted by the audio and visual encoders can be downloaded from the google drive link. After downloading, place the 'features' folder in the current directory to run experiments with our proposed method.

Citation

@inproceedings{wu2025avqacl,
  title={AVQACL: A Novel Benchmark for Audio-Visual Question Answering Continual Learning},
  author={Wu, Kaixuan and Li, Xinde and Li, Xinling and Hu, Chuanfei and Wu, Guoliang},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={3252--3261},
  year={2025}
}