Request for Pre-extracted Features of AV-Deepfake1M for UMMAFormer Reproduction

[123308]

Dear Author,
According to the UMMAFormer paper, extracting features for this dataset requires multiple complex preprocessing steps, including:

Frame extraction and optical flow computation using MMAction (v0.24.1)

Audio separation and feature extraction using BYOL-A

Integration of pre-trained TSN models for RGB and flow features

Given the complexity of this pipeline and the substantial computational resources required, I was wondering if you would be willing to share the pre-extracted features (both visual and audio) that you used in your experiments with the AV-Deepfake1M dataset.

This would greatly facilitate our reproduction efforts and help ensure the accuracy of our implementation. We would be happy to:

Sign any necessary data usage agreements

Provide proper attribution in our work

Share our reproduction results with you for verification

If the features cannot be shared publicly, we would be grateful for any guidance on the specific parameters and configurations you used in your feature extraction pipeline.

Thank you for considering this request. I greatly admire your work and look forward to the possibility of learning from your implementation.

[ControlNet]

Hi, thank you for your issue.

Given the complexity of this pipeline and the substantial computational resources required, I was wondering if you would be willing to share the pre-extracted features (both visual and audio) that you used in your experiments with the AV-Deepfake1M dataset.

Sorry considering these extracted frames and features consume terabytes which is even more than the dataset itself, we didn't keep them, but I know the official scripts work to extract the dataset features with minimal changes. Please follow the github guide of these features.

According to the UMMAFormer paper, extracting features for this dataset requires multiple complex preprocessing steps, including:
Frame extraction and optical flow computation using MMAction (v0.24.1)
Audio separation and feature extraction using BYOL-A
Integration of pre-trained TSN models for RGB and flow features

BTW, as we show in section 4.2 in our paper, we use the InternVideo as the video features for all baselines requiring video features including TriDet, ActionFormer, UMMAFormer. But of course, you can extract TSN features for these methods.

[123308]

Dear Authors,

Thank you very much for your impressive work on AV-Deepfake1M. The dataset and benchmark are extremely valuable for the community.

In the paper, you mention using InternVideo as the video feature extractor, but we could not find detailed descriptions of the exact pipeline, model variant, or implementation in the paper or the released code.

Could you kindly clarify:

1. Which InternVideo model was used (e.g., InternVideo1 / VideoMAE-based, ViT-B/L, pretrained on Kinetics or other datasets)?
2. How video features were extracted (clip length, sampling rate, pooling strategy such as mean pooling / CLS token)?
3. Whether the feature extraction code or configuration will be released, or if there is a reference implementation we could follow?

This information would greatly help reproducibility and benefit follow-up research using AV-Deepfake1M.

Thank you very much for your time and effort.

Best regards