{ "id": 136924, "title": "FAKE videos' frames", "url": "/competitions/deepfake-detection-challenge/discussion/136924", "author_name": "Schroter", "post_date": "2020-03-18T10:21:50.409000", "votes": 1, "comment_count": 14, "views": 0, "content": "

Hi @selimsef,

\n\n

You seem to be well versed in this DFDC challenge. Would you be able to let me know if all the frames in the train FAKE videos are deepfaked. i.e. if there are 100 frames in a videos that has been labeled as FAKE then all 100 frames have been deep faked?

\n\n

Furthermore, is it possible to find a repository of deepfaked images on the net that we could use for training?

\n\n

Also, I could not find the REAL videos for the sample given to run on the kaggle kernal. I get an error saying that the video could not be found.

\n\n

Thanks & Best Regards

\n\n

Michael

", "messages": [ { "id": 778370, "postDate": "2020-03-18T11:31:40.333Z", "content": "

These questions have been answered many times on the forums, but:

\n\n
    \n

  1. Not all the frames in fake videos have always been faked. Some of the fake videos have issues like that.

    2. \n

  2. See the External Data Disclosure Thread for datasets you can use.

    3. \n

  3. The sample training set does not contain real videos for all the fakes. You'll have to use the full dataset.

    4. \n
", "rawMarkdown": "These questions have been answered many times on the forums, but:\n\n1. Not all the frames in fake videos have always been faked. Some of the fake videos have issues like that.\n\n2. See the External Data Disclosure Thread for datasets you can use.\n\n3. The sample training set does not contain real videos for all the fakes. You'll have to use the full dataset.\n", "votes": 6, "replies": [ { "id": 778401, "postDate": "2020-03-18T12:12:29.247Z", "content": "

I dont agree with point 1, can be proved with simpicity.

", "rawMarkdown": "I dont agree with point 1, can be proved with simpicity." }, { "id": 778605, "postDate": "2020-03-18T15:31:16.137Z", "content": "

I should clarify: not all the faces in the fake videos have always been faked in every frame. But there usually is some other modification of the frame (a blurry blob flying around, for example).

", "rawMarkdown": "I should clarify: not all the _faces_ in the fake videos have always been faked in every frame. But there usually is some other modification of the frame (a blurry blob flying around, for example).", "votes": 2 }, { "id": 778660, "postDate": "2020-03-18T16:16:23.530Z", "content": "

I agree with point 1. There are some videos where there are no fake faces, and there is a flyiing blob somewhere in frame (due to detector failure). Hopefully private holdout will be cleaner as log loss metric is unforgiving and we cannot predict 1 when face is always real in a video. Otherwise it is just a lottery with big prizes.

", "rawMarkdown": "I agree with point 1. There are some videos where there are no fake faces, and there is a flyiing blob somewhere in frame (due to detector failure). Hopefully private holdout will be cleaner as log loss metric is unforgiving and we cannot predict 1 when face is always real in a video. Otherwise it is just a lottery with big prizes.", "votes": 7 }, { "id": 778785, "postDate": "2020-03-18T18:19:48.480Z", "content": "

I am saying the thing again, point 1 is wrong, not that it matters, but it is wrong. The case of flying blob is a very exceptional case, but still it does make the point 1 wrong, cuz if the point 1 is not wrong my data creation program could not have ran, the first version of my data creation had a thing just to highly focus on the data where fake do not have real faces even at a single frame, the program did ran through all the videos but the counter did not go even 1 up. There is not 1 possibility to this problem, where I can prove it that even faces are changing, at the same point, I can prove in a very special scenerio, like maybe in 1 or 2 videos, that faces are not changing, but because that scenerio is not possible to happen without we know exact real and fakes of public/private test set videos, hence it is not possible and the first case with \"Every face in the fake video is fake\" will be proven victorious, my theory does take in mind that flying blob. I dont want to share the massive insights I did just to find how can I clean the labels perfectly right now (FAILED), its way too information and stats to do public.

", "rawMarkdown": "I am saying the thing again, point 1 is wrong, not that it matters, but it is wrong. The case of flying blob is a very exceptional case, but still it does make the point 1 wrong, cuz if the point 1 is not wrong my data creation program could not have ran, the first version of my data creation had a thing just to highly focus on the data where fake do not have real faces even at a single frame, the program did ran through all the videos but the counter did not go even 1 up. There is not 1 possibility to this problem, where I can prove it that even faces are changing, at the same point, I can prove in a very special scenerio, like maybe in 1 or 2 videos, that faces are not changing, but because that scenerio is not possible to happen without we know exact real and fakes of public/private test set videos, hence it is not possible and the first case with \"Every face in the fake video is fake\" will be proven victorious, my theory does take in mind that flying blob. I dont want to share the massive insights I did just to find how can I clean the labels perfectly right now (FAILED), its way too information and stats to do public.", "votes": -2 }, { "id": 778836, "postDate": "2020-03-18T19:34:46.413Z", "content": "

Sorry, but I didn't understand any of that. Not saying you're wrong but I'm not really sure what it is that you're actually saying.

\n\n

I wrote a script to compare the real faces with the same crop region in the fake videos, and for a number of these fake videos there was no measurable difference with the real face. At least not measurable beyond any compression noise. (It is possible that these were all examples of the flying blob.)

", "rawMarkdown": "Sorry, but I didn't understand any of that. Not saying you're wrong but I'm not really sure what it _is_ that you're actually saying.\n\nI wrote a script to compare the real faces with the same crop region in the fake videos, and for a number of these fake videos there was no measurable difference with the real face. At least not measurable beyond any compression noise. (It is possible that these were all examples of the flying blob.)", "votes": 6 }, { "id": 778862, "postDate": "2020-03-18T20:09:34.853Z", "content": "

That is somewhat a correct assumption, the difference is there, in most of the cases it is so minor, that you can say like it is not even there and that is the reason it does not even matter.

", "rawMarkdown": "That is somewhat a correct assumption, the difference is there, in most of the cases it is so minor, that you can say like it is not even there and that is the reason it does not even matter." }, { "id": 778870, "postDate": "2020-03-18T20:17:27.567Z", "content": "

The difference exists due to compression even if faceswap was applied to a different region (on clouds for example) and face was not touched at all.

", "rawMarkdown": "The difference exists due to compression even if faceswap was applied to a different region (on clouds for example) and face was not touched at all.", "votes": 6 }, { "id": 778874, "postDate": "2020-03-18T20:23:37.187Z", "content": "

@selimsef \nThank You, that solves so many problems, and a new place for the data creation but its just not worth in the time of last 10 days.

", "rawMarkdown": "@selimsef \nThank You, that solves so many problems, and a new place for the data creation but its just not worth in the time of last 10 days." }, { "id": 779055, "postDate": "2020-03-19T01:34:06.890Z", "content": "

@humalag is it a valid assumption that only faces are fake? In some videos flagged as fake, the only thing that appears off, is the size of the person in relation to the rest of the room. I don't have the video name ready, will post it if I find it again

", "rawMarkdown": "@humalag is it a valid assumption that only faces are fake? In some videos flagged as fake, the only thing that appears off, is the size of the person in relation to the rest of the room. I don't have the video name ready, will post it if I find it again\n" }, { "id": 779239, "postDate": "2020-03-19T06:04:46.887Z", "content": "

I show in a different post how there are real face frames in fake videos. The difference can be measured with structural similarity to the extent that it is clearly just a difference in compression. There are also misaligned frames between real and fake videos that needs to be accounted for if you are relying on the same bounding boxes.

\n\n

https://www.kaggle.com/c/deepfake-detection-challenge/discussion/135648

", "rawMarkdown": "I show in a different post how there are real face frames in fake videos. The difference can be measured with structural similarity to the extent that it is clearly just a difference in compression. There are also misaligned frames between real and fake videos that needs to be accounted for if you are relying on the same bounding boxes.\n\nhttps://www.kaggle.com/c/deepfake-detection-challenge/discussion/135648", "votes": 1 }, { "id": 779472, "postDate": "2020-03-19T11:40:12.267Z", "content": "

It is natural to expect some fraction of the training examples to be incorrect in such a large dataset. It's also not necessarily a bad thing, as noisy data can help the model learn better.

", "rawMarkdown": "It is natural to expect some fraction of the training examples to be incorrect in such a large dataset. It's also not necessarily a bad thing, as noisy data can help the model learn better.", "votes": 2 } ] }, { "id": 778294, "postDate": "2020-03-18T10:21:50.410Z", "content": "

Hi @selimsef,

\n\n

You seem to be well versed in this DFDC challenge. Would you be able to let me know if all the frames in the train FAKE videos are deepfaked. i.e. if there are 100 frames in a videos that has been labeled as FAKE then all 100 frames have been deep faked?

\n\n

Furthermore, is it possible to find a repository of deepfaked images on the net that we could use for training?

\n\n

Also, I could not find the REAL videos for the sample given to run on the kaggle kernal. I get an error saying that the video could not be found.

\n\n

Thanks & Best Regards

\n\n

Michael

", "rawMarkdown": "Hi @selimsef,\n\nYou seem to be well versed in this DFDC challenge. Would you be able to let me know if all the *frames* in the train FAKE videos are deepfaked. i.e. if there are 100 frames in a videos that has been labeled as FAKE then all 100 frames have been deep faked?\n\nFurthermore, is it possible to find a repository of deepfaked images on the net that we could use for training? \n\nAlso, I could not find the REAL videos for the sample given to run on the kaggle kernal. I get an error saying that the video could not be found.\n\nThanks & Best Regards\n\nMichael", "votes": 1 }, { "id": 779246, "postDate": "2020-03-19T06:18:48.283Z", "content": "

Hi Guys,

\n\n

Thanks for all the help. Really appreciate it.

\n\n

Thanks & Best Regards

\n\n

Michael

", "rawMarkdown": "Hi Guys,\n\nThanks for all the help. Really appreciate it. \n\nThanks & Best Regards\n\nMichael" }, { "id": 778338, "postDate": "2020-03-18T11:01:41.967Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 778370, "author_name": "Human Analog", "author_url": "", "post_date": "2020-03-18T11:31:40.333000", "content": "

These questions have been answered many times on the forums, but:

\n\n
    \n

  1. Not all the frames in fake videos have always been faked. Some of the fake videos have issues like that.

    2. \n

  2. See the External Data Disclosure Thread for datasets you can use.

    3. \n

  3. The sample training set does not contain real videos for all the fakes. You'll have to use the full dataset.

    4. \n
", "votes": 6, "replies": [ { "id": 778401, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2020-03-18T12:12:29.247000", "content": "

I dont agree with point 1, can be proved with simpicity.

", "votes": 0, "replies": [] }, { "id": 778605, "author_name": "Human Analog", "author_url": "", "post_date": "2020-03-18T15:31:16.137000", "content": "

I should clarify: not all the faces in the fake videos have always been faked in every frame. But there usually is some other modification of the frame (a blurry blob flying around, for example).

", "votes": 2, "replies": [] }, { "id": 778660, "author_name": "Selim Seferbekov", "author_url": "", "post_date": "2020-03-18T16:16:23.530000", "content": "

I agree with point 1. There are some videos where there are no fake faces, and there is a flyiing blob somewhere in frame (due to detector failure). Hopefully private holdout will be cleaner as log loss metric is unforgiving and we cannot predict 1 when face is always real in a video. Otherwise it is just a lottery with big prizes.

", "votes": 7, "replies": [] }, { "id": 778785, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2020-03-18T18:19:48.480000", "content": "

I am saying the thing again, point 1 is wrong, not that it matters, but it is wrong. The case of flying blob is a very exceptional case, but still it does make the point 1 wrong, cuz if the point 1 is not wrong my data creation program could not have ran, the first version of my data creation had a thing just to highly focus on the data where fake do not have real faces even at a single frame, the program did ran through all the videos but the counter did not go even 1 up. There is not 1 possibility to this problem, where I can prove it that even faces are changing, at the same point, I can prove in a very special scenerio, like maybe in 1 or 2 videos, that faces are not changing, but because that scenerio is not possible to happen without we know exact real and fakes of public/private test set videos, hence it is not possible and the first case with \"Every face in the fake video is fake\" will be proven victorious, my theory does take in mind that flying blob. I dont want to share the massive insights I did just to find how can I clean the labels perfectly right now (FAILED), its way too information and stats to do public.

", "votes": -2, "replies": [] }, { "id": 778836, "author_name": "Human Analog", "author_url": "", "post_date": "2020-03-18T19:34:46.413000", "content": "

Sorry, but I didn't understand any of that. Not saying you're wrong but I'm not really sure what it is that you're actually saying.

\n\n

I wrote a script to compare the real faces with the same crop region in the fake videos, and for a number of these fake videos there was no measurable difference with the real face. At least not measurable beyond any compression noise. (It is possible that these were all examples of the flying blob.)

", "votes": 6, "replies": [] }, { "id": 778862, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2020-03-18T20:09:34.853000", "content": "

That is somewhat a correct assumption, the difference is there, in most of the cases it is so minor, that you can say like it is not even there and that is the reason it does not even matter.

", "votes": 0, "replies": [] }, { "id": 778870, "author_name": "Selim Seferbekov", "author_url": "", "post_date": "2020-03-18T20:17:27.567000", "content": "

The difference exists due to compression even if faceswap was applied to a different region (on clouds for example) and face was not touched at all.

", "votes": 6, "replies": [] }, { "id": 778874, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2020-03-18T20:23:37.187000", "content": "

@selimsef \nThank You, that solves so many problems, and a new place for the data creation but its just not worth in the time of last 10 days.

", "votes": 0, "replies": [] }, { "id": 779055, "author_name": "Rajaram G", "author_url": "", "post_date": "2020-03-19T01:34:06.890000", "content": "

@humalag is it a valid assumption that only faces are fake? In some videos flagged as fake, the only thing that appears off, is the size of the person in relation to the rest of the room. I don't have the video name ready, will post it if I find it again

", "votes": 0, "replies": [] }, { "id": 779239, "author_name": "maralski", "author_url": "", "post_date": "2020-03-19T06:04:46.887000", "content": "

I show in a different post how there are real face frames in fake videos. The difference can be measured with structural similarity to the extent that it is clearly just a difference in compression. There are also misaligned frames between real and fake videos that needs to be accounted for if you are relying on the same bounding boxes.

\n\n

https://www.kaggle.com/c/deepfake-detection-challenge/discussion/135648

", "votes": 1, "replies": [] }, { "id": 779472, "author_name": "Human Analog", "author_url": "", "post_date": "2020-03-19T11:40:12.267000", "content": "

It is natural to expect some fraction of the training examples to be incorrect in such a large dataset. It's also not necessarily a bad thing, as noisy data can help the model learn better.

", "votes": 2, "replies": [] } ] }, { "id": 779246, "author_name": "Schroter", "author_url": "", "post_date": "2020-03-19T06:18:48.283000", "content": "

Hi Guys,

\n\n

Thanks for all the help. Really appreciate it.

\n\n

Thanks & Best Regards

\n\n

Michael

", "votes": 0, "replies": [] }, { "id": 778338, "author_name": "", "author_url": "", "post_date": "2020-03-18T11:01:41.967000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "778370": "These questions have been answered many times on the forums, but:\n\n1. Not all the frames in fake videos have always been faked. Some of the fake videos have issues like that.\n\n2. See the External Data Disclosure Thread for datasets you can use.\n\n3. The sample training set does not contain real videos for all the fakes. You'll have to use the full dataset.\n", "778294": "Hi @selimsef,\n\nYou seem to be well versed in this DFDC challenge. Would you be able to let me know if all the *frames* in the train FAKE videos are deepfaked. i.e. if there are 100 frames in a videos that has been labeled as FAKE then all 100 frames have been deep faked?\n\nFurthermore, is it possible to find a repository of deepfaked images on the net that we could use for training? \n\nAlso, I could not find the REAL videos for the sample given to run on the kaggle kernal. I get an error saying that the video could not be found.\n\nThanks & Best Regards\n\nMichael", "779246": "Hi Guys,\n\nThanks for all the help. Really appreciate it. \n\nThanks & Best Regards\n\nMichael", "778338": "" } }