search

Daisy-Zhang / Awesome-Deepfakes-Detection

A list of tools, papers and code related to Deepfake Detection.
MIT License
925 stars 91 forks source link

Questions about model testing #20

Open kk3lamb opened 4 weeks ago

kk3lamb commented 4 weeks ago

Thank you for your excellent work! How is the model for video-level forgery detection tested? Do you test all the frames of the test video? I see a lot of papers that do video forgery detection where the model input is a clip of several consecutive frames (e.g., 8 frames), so how is this tested? Is it averaging?

Daisy-Zhang commented 3 weeks ago

Hi~ For the video-level testing, usually the input is a frame clip and the model output is the one prediction. There are also some methods as you mentioned that take multiple frames and output multiple predictions, and in this case they usually calculate the averaging prediction as final score.

kk3lamb commented 3 weeks ago

  1. I have seen papers that say they want a fair comparison with image level models, some say they will bring all the frames of a test video to test, what is the average prediction you are talking about in such a case? If my model accepts 8 frames of clips, then a test video will have a very large number of clips, each of which will have a prediction, how do I determine if the video is predicted accurately? How should I take the average?

  2. Or can you tell me how the mainstream testing process works? Is it possible to take 32 frames of a test video?

Thank you very much for your answer, I'm new to forgery detection and your answer helps me a lot.