![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F806279%2F22925afa7fe24bd83a5d14b8487a008e%2F1frame.png?generation=1582643464876182&alt=media\")
I noticed that in some fake dialog videos, the face swap algorithm targets different interlocutors within one video. \nFor instance, abjgxazkfk.mp4 (part_2).
\n\n~1st second [left - Fake, right - Real] \n
~5th second [left - Real, right - Real]\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F806279%2Fa3984daa927ba6a39a4ae20ea1504fc9%2F2frame.png?generation=1582643466372967&alt=media\")
~6th second [left - Real, right - Fake]\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F806279%2F62cd95807b0ae0b62d25370c5a8d97d9%2F3frame.png?generation=1582643464816283&alt=media\")
It could be critical for labeling.
", "messages": [ { "id": "756290", "postDate": "02/25/2020 15:30:43", "content": "I noticed that in some fake dialog videos, the face swap algorithm targets different interlocutors within one video. \nFor instance, abjgxazkfk.mp4 (part_2).
\n\n~1st second [left - Fake, right - Real] \n
~5th second [left - Real, right - Real]\n
~6th second [left - Real, right - Fake]\n
It could be critical for labeling.
", "rawMarkdown": "I noticed that in some fake dialog videos, the face swap algorithm targets different interlocutors within one video. \nFor instance, abjgxazkfk.mp4 (part_2).\n\n~1st second [left - Fake, right - Real] \n\n\n~5th second [left - Real, right - Real]\n\n\n~6th second [left - Real, right - Fake]\n\n\n\n\nIt could be critical for labeling.", "votes": null }, { "id": "756340", "postDate": "02/25/2020 16:23:56", "content": "Note that there is a blur flying from one person to the other in this video. I believe this is an error mode in the deep fake generation program they used: it did not properly detect the faces and got confused.
", "rawMarkdown": "Note that there is a blur flying from one person to the other in this video. I believe this is an error mode in the deep fake generation program they used: it did not properly detect the faces and got confused.", "votes": null }, { "id": "756352", "postDate": "02/25/2020 16:38:40", "content": "Actually, I agree with your assumption. I've just realized, that in the inference mode, given this video, the probability of fakeness will be lower than it should be. Even with a perfect classifier.
", "rawMarkdown": "Actually, I agree with your assumption. I've just realized, that in the inference mode, given this video, the probability of fakeness will be lower than it should be. Even with a perfect classifier.", "votes": null }, { "id": "756368", "postDate": "02/25/2020 16:59:51", "content": "But as soon as you detect with enough confidence that one frame is fake, you can predict 1.0 for this video. ;-)
", "rawMarkdown": "But as soon as you detect with enough confidence that one frame is fake, you can predict 1.0 for this video. ;-)", "votes": null }, { "id": "756776", "postDate": "02/26/2020 04:07:52", "content": "This is a good point! Thanks.
", "rawMarkdown": "This is a good point! Thanks.", "votes": null }, { "id": "756878", "postDate": "02/26/2020 06:56:41", "content": "And you probably want a different scoring approach depending on whether there is one or multiple faces in the video.
", "rawMarkdown": "And you probably want a different scoring approach depending on whether there is one or multiple faces in the video.", "votes": null }, { "id": "759328", "postDate": "02/28/2020 22:31:37", "content": "some videos are even worse, the blur only appears at the end of the video and it wasn't even then on the face .. but i'd say most fakes are ok for a face based classifier but take more than one frame in inferring
", "rawMarkdown": "some videos are even worse, the blur only appears at the end of the video and it wasn't even then on the face .. but i'd say most fakes are ok for a face based classifier but take more than one frame in inferring", "votes": null }, { "id": "759427", "postDate": "02/29/2020 02:47:44", "content": "\"Judging based on face\" would inherently miss these videos. And this isn't just a \"one off\" situation, its pretty common. There are a few techniques they use for mutations and some of the techniques deal with video frame manipulation or audio manipulation instead of face manipulation.
", "rawMarkdown": "\"Judging based on face\" would inherently miss these videos. And this isn't just a \"one off\" situation, its pretty common. There are a few techniques they use for mutations and some of the techniques deal with video frame manipulation or audio manipulation instead of face manipulation.", "votes": null } ], "comments": [ { "id": 756340, "author_name": "humananalog", "author_url": "", "post_date": "02/25/2020 16:23:56", "content": "Note that there is a blur flying from one person to the other in this video. I believe this is an error mode in the deep fake generation program they used: it did not properly detect the faces and got confused.
", "votes": null, "replies": [ { "id": 756352, "author_name": "ruslankhalitov", "author_url": "", "post_date": "02/25/2020 16:38:40", "content": "Actually, I agree with your assumption. I've just realized, that in the inference mode, given this video, the probability of fakeness will be lower than it should be. Even with a perfect classifier.
", "votes": null, "replies": [] }, { "id": 756368, "author_name": "humananalog", "author_url": "", "post_date": "02/25/2020 16:59:51", "content": "But as soon as you detect with enough confidence that one frame is fake, you can predict 1.0 for this video. ;-)
", "votes": null, "replies": [] }, { "id": 756776, "author_name": "cooleel", "author_url": "", "post_date": "02/26/2020 04:07:52", "content": "This is a good point! Thanks.
", "votes": null, "replies": [] }, { "id": 756878, "author_name": "maralski", "author_url": "", "post_date": "02/26/2020 06:56:41", "content": "And you probably want a different scoring approach depending on whether there is one or multiple faces in the video.
", "votes": null, "replies": [] } ] }, { "id": 759328, "author_name": "basharallabadi", "author_url": "", "post_date": "02/28/2020 22:31:37", "content": "some videos are even worse, the blur only appears at the end of the video and it wasn't even then on the face .. but i'd say most fakes are ok for a face based classifier but take more than one frame in inferring
", "votes": null, "replies": [] }, { "id": 759427, "author_name": "quiteconfused", "author_url": "", "post_date": "02/29/2020 02:47:44", "content": "\"Judging based on face\" would inherently miss these videos. And this isn't just a \"one off\" situation, its pretty common. There are a few techniques they use for mutations and some of the techniques deal with video frame manipulation or audio manipulation instead of face manipulation.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "756290": "I noticed that in some fake dialog videos, the face swap algorithm targets different interlocutors within one video. \nFor instance, abjgxazkfk.mp4 (part_2).\n\n~1st second [left - Fake, right - Real] \n\n\n~5th second [left - Real, right - Real]\n\n\n~6th second [left - Real, right - Fake]\n\n\n\n\nIt could be critical for labeling.", "756340": "Note that there is a blur flying from one person to the other in this video. I believe this is an error mode in the deep fake generation program they used: it did not properly detect the faces and got confused.", "756352": "Actually, I agree with your assumption. I've just realized, that in the inference mode, given this video, the probability of fakeness will be lower than it should be. Even with a perfect classifier.", "756368": "But as soon as you detect with enough confidence that one frame is fake, you can predict 1.0 for this video. ;-)", "756776": "This is a good point! Thanks.", "756878": "And you probably want a different scoring approach depending on whether there is one or multiple faces in the video.", "759328": "some videos are even worse, the blur only appears at the end of the video and it wasn't even then on the face .. but i'd say most fakes are ok for a face based classifier but take more than one frame in inferring", "759427": "\"Judging based on face\" would inherently miss these videos. And this isn't just a \"one off\" situation, its pretty common. There are a few techniques they use for mutations and some of the techniques deal with video frame manipulation or audio manipulation instead of face manipulation." }, "source": "meta" }