Accuracy of provided trained models does not match reported results in the paper about faceforensics HOT 4 OPEN

I have a simliar problem. I test the model face_detection/xception/all_raw.p provided in the paper on the raw videos. The accuracy is high on the DF, F2F and FS, while it's only about 0.75 acc on the NeuralTextures.
I just want to know if there is any point I missed when test on NT?

from faceforensics.

I have a simliar problem. I test the model face_detection/xception/all_raw.p provided in the paper on the raw videos. The accuracy is high on the DF, F2F and FS, while it's only about 0.75 acc on the NeuralTextures.
I just want to know if there is any point I missed when test on NT?

May I ask how the acc value is calculated after your test? Is it the average of the sum of the frames？

from faceforensics.

I have a simliar problem. I test the model face_detection/xception/all_raw.p provided in the paper on the raw videos. The accuracy is high on the DF, F2F and FS, while it's only about 0.75 acc on the NeuralTextures.
I just want to know if there is any point I missed when test on NT?

May I ask how the acc value is calculated after your test? Is it the average of the sum of the frames？

I calculated the acc of each video using the predicted value of all frames in this video. Then I averaged the acc of all videos.

from faceforensics.

I have a simliar problem. I test the model face_detection/xception/all_raw.p provided in the paper on the raw videos. The accuracy is high on the DF, F2F and FS, while it's only about 0.75 acc on the NeuralTextures.
I just want to know if there is any point I missed when test on NT?

May I ask how the acc value is calculated after your test? Is it the average of the sum of the frames？

I calculated the acc of each video using the predicted value of all frames in this video. Then I averaged the acc of all videos.

I have a simliar problem. I test the model face_detection/xception/all_raw.p provided in the paper on the raw videos. The accuracy is high on the DF, F2F and FS, while it's only about 0.75 acc on the NeuralTextures.
I just want to know if there is any point I missed when test on NT?

May I ask how the acc value is calculated after your test? Is it the average of the sum of the frames？

I calculated the acc of each video using the predicted value of all frames in this video. Then I averaged the acc of all videos.

同学你好，我有些简单的问题想请问你，
作者的训练集是提取视频的所有帧，还是选取第一帧或者随机提取一帧的？
在训练完成后，文章里的测试结果是如何计算的，是对整个测试集的视频测试，然后统计这个视频每帧的分类结果多数的值，作为该视频的预测，还是也是跟训练一样，取一帧测试，又或者是将所有帧提取出来计算全部的分类结果？
谢谢！

from faceforensics.