Question about the AVQA dataset

[tanganke]

Thank you for your great work!

I am trying to run the AVQA evaluation and I downloaded AVQA from here.
The folder structure is as follows:

.
├── AVQA
│   ├── AVQA_dataset
│   │   ├── avqa_download_urls.csv
│   │   ├── train_qa.json
│   │   └── val_qa.json
│   └── AVQA_extracted_features
│       ├── AVQA_appearance_resnet101_feat.h5
│       ├── AVQA_audio_PANNs_feat.h5
│       └── AVQA_motion_resnext101_feat.h5
└── AVQA.zip

But when I run the script, it raises FileNotFoundErrors such as
No such file or directory: '.../ModelCompose/data/evaluation_datasets/AVQA/image/EdhHwpq1Isg_000064.jpg'

How can I get those data files?

[threegold116]

You can find origin files of avqa in Hoar012/avqa_data (huggingface hub name).

[WalkerWorldPeace]

You can find origin files of avqa in Hoar012/avqa_data (huggingface hub name).

Hello, I was unable to find the Hoar012/avqa_data on HuggingFace. Could you kindly provide further clarification? Thank you!

[threegold116]

You can find origin files of avqa in Hoar012/avqa_data (huggingface hub name).

Hello, I was unable to find the Hoar012/avqa_data on HuggingFace. Could you kindly provide further clarification? Thank you!

AVQA's origin files are from VGGSound. You can download origin videos from Loie/VGGSound and Nano1337/vggsound_08.

[WalkerWorldPeace]

Since I noticed that the AVQA dataset only provides extracted features and an avqa_download_csv file, which only contains the starting seconds of the videos, how can I download and process the corresponding videos into the AVQA format based on this CSV file?

[kaykobad]

I think I figured it out. We can get the Audios and Videos of AVQA from this link: Huggingface Repo

I am still unsure about the images part.

FYI: @WalkerWorldPeace @tanganke