{
  "id": 131842,
  "title": "Has anyone done an analysis on how many real frames are found in a labelled 'FAKE' video?",
  "url": "/competitions/deepfake-detection-challenge/discussion/131842",
  "author_name": "",
  "post_date": "2020-02-22T04:19:50.752051300Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "",
  "messages": [
    {
      "id": "753357",
      "postDate": "02/22/2020 04:19:50",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "754756",
      "postDate": "02/24/2020 02:41:42",
      "content": "<p>I'm working on that. \nI just compare the fake video with its corresponding real video frame by frame.\nWhat have u found?</p>",
      "rawMarkdown": "I'm working on that. \nI just compare the fake video with its corresponding real video frame by frame.\nWhat have u found?",
      "votes": null
    },
    {
      "id": "763900",
      "postDate": "03/05/2020 00:36:16",
      "content": "<p>found that there is a difference in each corresponding frame for real and fake videos in almost all frames, may be doing the difference incorrectly. have you found the same thing?</p>",
      "rawMarkdown": "found that there is a difference in each corresponding frame for real and fake videos in almost all frames, may be doing the difference incorrectly. have you found the same thing?",
      "votes": null
    },
    {
      "id": "764321",
      "postDate": "03/05/2020 10:44:01",
      "content": "<p>A lot of these differences come from the MP4 encoding and are really nothing more than just noise. So if you want to find the true difference, you'll have to compensate for this noise first.</p>",
      "rawMarkdown": "A lot of these differences come from the MP4 encoding and are really nothing more than just noise. So if you want to find the true difference, you'll have to compensate for this noise first.",
      "votes": null
    },
    {
      "id": "764391",
      "postDate": "03/05/2020 12:14:07",
      "content": "<p>How do you recommend doing this? And in your experience does this seem useful to do for getting good results?</p>",
      "rawMarkdown": "How do you recommend doing this? And in your experience does this seem useful to do for getting good results?",
      "votes": null
    },
    {
      "id": "764468",
      "postDate": "03/05/2020 13:53:47",
      "content": "<p>I used a combination of SSIM, MSE, and L2 distance between the face crops from the real and fake videos (where the fake crops were taking in the exact same location as the real crops). </p>\n\n<p>This approach works but is not ideal because some fake faces really are fakes but the difference is very subtle, and they would be considered the same by my approach.</p>\n\n<p>Using this method I removed about 10% of my training data (i.e. fake face crops that are too similar to the real face, but it also catches most false positives that are not really faces) but my model actually became a lot worse... /shrug</p>",
      "rawMarkdown": "I used a combination of SSIM, MSE, and L2 distance between the face crops from the real and fake videos (where the fake crops were taking in the exact same location as the real crops). \n\nThis approach works but is not ideal because some fake faces really are fakes but the difference is very subtle, and they would be considered the same by my approach.\n\nUsing this method I removed about 10% of my training data (i.e. fake face crops that are too similar to the real face, but it also catches most false positives that are not really faces) but my model actually became a lot worse... /shrug",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 754756,
      "author_name": "zhanglonghao",
      "author_url": "",
      "post_date": "02/24/2020 02:41:42",
      "content": "<p>I'm working on that. \nI just compare the fake video with its corresponding real video frame by frame.\nWhat have u found?</p>",
      "votes": null,
      "replies": [
        {
          "id": 763900,
          "author_name": "harisanthanam",
          "author_url": "",
          "post_date": "03/05/2020 00:36:16",
          "content": "<p>found that there is a difference in each corresponding frame for real and fake videos in almost all frames, may be doing the difference incorrectly. have you found the same thing?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 764321,
          "author_name": "humananalog",
          "author_url": "",
          "post_date": "03/05/2020 10:44:01",
          "content": "<p>A lot of these differences come from the MP4 encoding and are really nothing more than just noise. So if you want to find the true difference, you'll have to compensate for this noise first.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 764391,
          "author_name": "harisanthanam",
          "author_url": "",
          "post_date": "03/05/2020 12:14:07",
          "content": "<p>How do you recommend doing this? And in your experience does this seem useful to do for getting good results?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 764468,
          "author_name": "humananalog",
          "author_url": "",
          "post_date": "03/05/2020 13:53:47",
          "content": "<p>I used a combination of SSIM, MSE, and L2 distance between the face crops from the real and fake videos (where the fake crops were taking in the exact same location as the real crops). </p>\n\n<p>This approach works but is not ideal because some fake faces really are fakes but the difference is very subtle, and they would be considered the same by my approach.</p>\n\n<p>Using this method I removed about 10% of my training data (i.e. fake face crops that are too similar to the real face, but it also catches most false positives that are not really faces) but my model actually became a lot worse... /shrug</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "753357": "",
    "754756": "I'm working on that. \nI just compare the fake video with its corresponding real video frame by frame.\nWhat have u found?",
    "763900": "found that there is a difference in each corresponding frame for real and fake videos in almost all frames, may be doing the difference incorrectly. have you found the same thing?",
    "764321": "A lot of these differences come from the MP4 encoding and are really nothing more than just noise. So if you want to find the true difference, you'll have to compensate for this noise first.",
    "764391": "How do you recommend doing this? And in your experience does this seem useful to do for getting good results?",
    "764468": "I used a combination of SSIM, MSE, and L2 distance between the face crops from the real and fake videos (where the fake crops were taking in the exact same location as the real crops). \n\nThis approach works but is not ideal because some fake faces really are fakes but the difference is very subtle, and they would be considered the same by my approach.\n\nUsing this method I removed about 10% of my training data (i.e. fake face crops that are too similar to the real face, but it also catches most false positives that are not really faces) but my model actually became a lot worse... /shrug"
  },
  "source": "meta"
}