{
  "id": 136029,
  "title": "Curious, what has been faked? Is actor different?",
  "url": "/competitions/deepfake-detection-challenge/discussion/136029",
  "author_name": "",
  "post_date": "2020-03-17T05:17:12.715456700Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi,</p>\n\n<p>For some videos, I do not understand what has been faked? Real and faked video they both seem to be same (there is some artifacts in fake), but I am not sure what was the intention? Actor and video and audio is actually same.</p>\n\n<p>Is actor different in this? from: \"/kaggle/input/deepfake-detection-challenge/train_sample_videos/\" for convenience\nFake: abofeumbvv.mp4       Real counterpart: atvmxvwyns.mp4</p>",
  "messages": [
    {
      "id": "776059",
      "postDate": "03/17/2020 05:17:12",
      "content": "<p>Hi,</p>\n\n<p>For some videos, I do not understand what has been faked? Real and faked video they both seem to be same (there is some artifacts in fake), but I am not sure what was the intention? Actor and video and audio is actually same.</p>\n\n<p>Is actor different in this? from: \"/kaggle/input/deepfake-detection-challenge/train_sample_videos/\" for convenience\nFake: abofeumbvv.mp4       Real counterpart: atvmxvwyns.mp4</p>",
      "rawMarkdown": "Hi,\n\nFor some videos, I do not understand what has been faked? Real and faked video they both seem to be same (there is some artifacts in fake), but I am not sure what was the intention? Actor and video and audio is actually same.\n\nIs actor different in this? from: \"/kaggle/input/deepfake-detection-challenge/train_sample_videos/\" for convenience\nFake: abofeumbvv.mp4       Real counterpart: atvmxvwyns.mp4",
      "votes": null
    },
    {
      "id": "780375",
      "postDate": "03/20/2020 07:46:12",
      "content": "<p>They might have manually changed one pixel from 10 to 11.</p>",
      "rawMarkdown": "They might have manually changed one pixel from 10 to 11.",
      "votes": null
    },
    {
      "id": "780659",
      "postDate": "03/20/2020 13:46:36",
      "content": "<p><a href=\"/aknirala\">@aknirala</a> I think this has to do with how the data was created. This was an automated process where they run some different deepfake algorithms on top of the real videos, and labeled the output of those as fake.\nThese algorithms work, by first performing face detection, and then alter the faces. Now, face detection is not a simple task in every video, sometimes the algorithm gets it wrong and detect faces somewhere where no faces are present, and tries to apply the same algorithm there.\nSomewhere in another topic there is a nice example of a face being placed in the sky on top of some clouds .</p>\n\n<p>This might explain what you are seeing. Regarding the training of your model with this data, it is fine to do so, as long as the wrongly labeled data remains a very small percentage. Of course if you'd know all wrongly labelled videos, it would be wise to remove them, but this is just a normal thing when training models, there is always a percentage of data that is wrongly labelled</p>",
      "rawMarkdown": "aknirala I think this has to do with how the data was created. This was an automated process where they run some different deepfake algorithms on top of the real videos, and labeled the output of those as fake.\nThese algorithms work, by first performing face detection, and then alter the faces. Now, face detection is not a simple task in every video, sometimes the algorithm gets it wrong and detect faces somewhere where no faces are present, and tries to apply the same algorithm there.\nSomewhere in another topic there is a nice example of a face being placed in the sky on top of some clouds .\n\nThis might explain what you are seeing. Regarding the training of your model with this data, it is fine to do so, as long as the wrongly labeled data remains a very small percentage. Of course if you'd know all wrongly labelled videos, it would be wise to remove them, but this is just a normal thing when training models, there is always a percentage of data that is wrongly labelled",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 780375,
      "author_name": "fuzhuolin",
      "author_url": "",
      "post_date": "03/20/2020 07:46:12",
      "content": "<p>They might have manually changed one pixel from 10 to 11.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 780659,
      "author_name": "ngcferreira",
      "author_url": "",
      "post_date": "03/20/2020 13:46:36",
      "content": "<p><a href=\"/aknirala\">@aknirala</a> I think this has to do with how the data was created. This was an automated process where they run some different deepfake algorithms on top of the real videos, and labeled the output of those as fake.\nThese algorithms work, by first performing face detection, and then alter the faces. Now, face detection is not a simple task in every video, sometimes the algorithm gets it wrong and detect faces somewhere where no faces are present, and tries to apply the same algorithm there.\nSomewhere in another topic there is a nice example of a face being placed in the sky on top of some clouds .</p>\n\n<p>This might explain what you are seeing. Regarding the training of your model with this data, it is fine to do so, as long as the wrongly labeled data remains a very small percentage. Of course if you'd know all wrongly labelled videos, it would be wise to remove them, but this is just a normal thing when training models, there is always a percentage of data that is wrongly labelled</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "776059": "Hi,\n\nFor some videos, I do not understand what has been faked? Real and faked video they both seem to be same (there is some artifacts in fake), but I am not sure what was the intention? Actor and video and audio is actually same.\n\nIs actor different in this? from: \"/kaggle/input/deepfake-detection-challenge/train_sample_videos/\" for convenience\nFake: abofeumbvv.mp4       Real counterpart: atvmxvwyns.mp4",
    "780375": "They might have manually changed one pixel from 10 to 11.",
    "780659": "aknirala I think this has to do with how the data was created. This was an automated process where they run some different deepfake algorithms on top of the real videos, and labeled the output of those as fake.\nThese algorithms work, by first performing face detection, and then alter the faces. Now, face detection is not a simple task in every video, sometimes the algorithm gets it wrong and detect faces somewhere where no faces are present, and tries to apply the same algorithm there.\nSomewhere in another topic there is a nice example of a face being placed in the sky on top of some clouds .\n\nThis might explain what you are seeing. Regarding the training of your model with this data, it is fine to do so, as long as the wrongly labeled data remains a very small percentage. Of course if you'd know all wrongly labelled videos, it would be wise to remove them, but this is just a normal thing when training models, there is always a percentage of data that is wrongly labelled"
  },
  "source": "meta"
}