Datasets:
Videos
Hello, are you planning to release videos in addition to the extracted frames? For instance, one might want to perform a temporal analysis of deepfake videos, which is impossible to do based only on specifically extracted frames.
Hi, we currently do not plan to do this. If the community has strong demand and we can spare more time on this, we may do it. Currently, I am super busy. But, please rest assured that the videos are stored safely in our disk.
I think unprocessed videos are much more valuable for the community because:
Any processed data uses outdated methods for preprocessing (e.g., detector) or normalization/alignment techniques that might differ between deepfake detection methods. Making the method rely on specific, sometimes unnecessary priors.
There is a growing trend in the community towards video-based detection and spatio-temporal and multimodal approaches, which require access to full video sequences.
I think contributing this data will be invaluable.
I agreed. Mark this as TODO.