It seems that on quite a few of the training videos, the method used to create the fake didn't work properly.
\n\nFor example, see acazlolrpz.mp4. You can see a blurry shape flying between the faces of these two people. My guess is that the particular deepfake method used had a problem detecting the faces in this video and got confused.
\n\nLet's say in frame 0 it detected the face of person A, in frames 1-9 it doesn't detect anything, and in frame 10 it detects the face of person B. Now it thinks the person moved from A's position to B's position and in frames 1 - 9 it will interpolate between these positions -- but since that's not what's really happening (it's not one person moving but two different people), this ends up blending the fake face with the background.
\n\nThis particular thing appears to happen on quite a number of the videos in the training set (also when there's just one person in the video). There are other examples of really bad-looking results too.
\n\nNo one will be fooled by a video like this, so it seems a bit silly that they appear in large amounts in this dataset.
\n\nI'm half tempted to build a classifier that can detect whether the video contains weird looking artifacts, in which case it almost certainly is fake. That would help getting a good score on the leaderboard (assuming this same thing happens in the test set) but it doesn't really help us detect deepfakes -- only really bad ones that don't fool anyone.
", "messages": [ { "id": 711315, "postDate": "2020-01-05T23:23:01.330Z", "content": "It seems that on quite a few of the training videos, the method used to create the fake didn't work properly.
\n\nFor example, see acazlolrpz.mp4. You can see a blurry shape flying between the faces of these two people. My guess is that the particular deepfake method used had a problem detecting the faces in this video and got confused.
\n\nLet's say in frame 0 it detected the face of person A, in frames 1-9 it doesn't detect anything, and in frame 10 it detects the face of person B. Now it thinks the person moved from A's position to B's position and in frames 1 - 9 it will interpolate between these positions -- but since that's not what's really happening (it's not one person moving but two different people), this ends up blending the fake face with the background.
\n\nThis particular thing appears to happen on quite a number of the videos in the training set (also when there's just one person in the video). There are other examples of really bad-looking results too.
\n\nNo one will be fooled by a video like this, so it seems a bit silly that they appear in large amounts in this dataset.
\n\nI'm half tempted to build a classifier that can detect whether the video contains weird looking artifacts, in which case it almost certainly is fake. That would help getting a good score on the leaderboard (assuming this same thing happens in the test set) but it doesn't really help us detect deepfakes -- only really bad ones that don't fool anyone.
", "rawMarkdown": "It seems that on quite a few of the training videos, the method used to create the fake didn't work properly. \n\nFor example, see **acazlolrpz.mp4**. You can see a blurry shape flying between the faces of these two people. My guess is that the particular deepfake method used had a problem detecting the faces in this video and got confused. \n\nLet's say in frame 0 it detected the face of person A, in frames 1-9 it doesn't detect anything, and in frame 10 it detects the face of person B. Now it thinks the person moved from A's position to B's position and in frames 1 - 9 it will interpolate between these positions -- but since that's not what's really happening (it's not one person moving but two different people), this ends up blending the fake face with the background.\n\nThis particular thing appears to happen on quite a number of the videos in the training set (also when there's just one person in the video). There are other examples of really bad-looking results too. \n\nNo one will be fooled by a video like this, so it seems a bit silly that they appear in large amounts in this dataset.\n\nI'm half tempted to build a classifier that can detect whether the video contains weird looking artifacts, in which case it almost certainly is fake. That would help getting a good score on the leaderboard (assuming this same thing happens in the test set) but it doesn't really help us detect deepfakes -- only really bad ones that don't fool anyone.", "votes": 4 } ], "comments": [], "raw_markdown_by_id": { "711315": "It seems that on quite a few of the training videos, the method used to create the fake didn't work properly. \n\nFor example, see **acazlolrpz.mp4**. You can see a blurry shape flying between the faces of these two people. My guess is that the particular deepfake method used had a problem detecting the faces in this video and got confused. \n\nLet's say in frame 0 it detected the face of person A, in frames 1-9 it doesn't detect anything, and in frame 10 it detects the face of person B. Now it thinks the person moved from A's position to B's position and in frames 1 - 9 it will interpolate between these positions -- but since that's not what's really happening (it's not one person moving but two different people), this ends up blending the fake face with the background.\n\nThis particular thing appears to happen on quite a number of the videos in the training set (also when there's just one person in the video). There are other examples of really bad-looking results too. \n\nNo one will be fooled by a video like this, so it seems a bit silly that they appear in large amounts in this dataset.\n\nI'm half tempted to build a classifier that can detect whether the video contains weird looking artifacts, in which case it almost certainly is fake. That would help getting a good score on the leaderboard (assuming this same thing happens in the test set) but it doesn't really help us detect deepfakes -- only really bad ones that don't fool anyone." } }