some questions about this dataset

[mowen9]

Thank you for sharing the code and dataset.
I have two questions to ask you for help.

1. how to split this dataset into Training and Test sets? I do not find the Split File.

2. Some descriptions are missing. Can you give a detailed description of [data path for video] and [data path for audio]? How to obtain these two folders. Can you provide the data preprocessor?

python triain_main.py -v=<data path for video> -a=<data path for audio>

Your prompt reply would be greatly appreciated.

[yinghuozijin]

Thank you for sharing the code and dataset. I have two questions to ask you for help.

1. how to split this dataset into Training and Test sets? I do not find the Split File.
2. Some descriptions are missing. Can you give a detailed description of [data path for video] and [data path for audio]? How to obtain these two folders. Can you provide the data preprocessor?

python triain_main.py -v=<data path for video> -a=<data path for audio>

Your prompt reply would be greatly appreciated.

About the second question ：
The video files include audio stream.So the paths of both are same.

[hasamkhalid]

I hope the questions and answers to the questions will be helpful to you:

Q1: how did you split the data in the train test split?
Q2: how do you deal with the Data Imbalance issue while training the models?

There are two ways of dealing with this data imbalance issue. I will explain both below:

I. To run experiments in both of our papers, we performed frame-level detection.
The number of frames extracted from Real videos was more than the number of frames extracted from Fake videos (Category B, C, D).
For example, we have 500 Real videos and around 10,000 videos in Category C and D.
So, we extracted 120 frames from Real videos : 120 x 500 = 60,000 Frames.
And we extracted 6 frames from Fake videos : 6 x 10,000 = 60,000 Frames.
By extracting frames this way, we got rid of the data imbalance issue.
Moreover, we have not provided any split file with the dataset. So you will not be able to find it.
But, we are working on it and we may provide train/test split once we decide.

II. There is another way of coping with the data imbalance issue that we briefly mentioned in the papers as well.
Since we collected Real videos from the VoxCeleb-2 dataset, the VoxCeleb dataset contains around a million Real videos.
As we have fewer Real videos in our FakeAVCeleb dataset, to get rid of the data imbalance issue we can get more
Real videos from the VoxCeleb dataset to train our models and run our experiments.
Also, we have provided the actual ids of the persons that we selected from the VoxCeleb dataset and selected a single video per person,
we can easily select more videos of the same person from the VoxCeleb dataset.