{
  "id": 135453,
  "title": "The way to calculate the percentage of prediction?",
  "url": "/competitions/deepfake-detection-challenge/discussion/135453",
  "author_name": "",
  "post_date": "2020-03-14T00:23:53.352690700Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>DDC provided a way to calculate the deep fake result percentage? \nif 1/300 frames are fake Do we determine this is fake or real? \nor just 1/300% as fake? </p>",
  "messages": [
    {
      "id": "771264",
      "postDate": "03/14/2020 00:23:53",
      "content": "<p>DDC provided a way to calculate the deep fake result percentage? \nif 1/300 frames are fake Do we determine this is fake or real? \nor just 1/300% as fake? </p>",
      "rawMarkdown": "DDC provided a way to calculate the deep fake result percentage? \nif 1/300 frames are fake Do we determine this is fake or real? \nor just 1/300% as fake?",
      "votes": null
    },
    {
      "id": "771313",
      "postDate": "03/14/2020 02:22:45",
      "content": "<p>I am just using a statistical metric (e.g. mean etc.) from all the prediction probabilities. It is one of the biggest challenges  I am facing, as no matter what the metric is, if your case happens (1/300 is FAKE), it is going to cause a big log loss. Again, my model is also not that confident to say, \"yes, I found a FAKE frame, so you call the video FAKE!\". Life is tough!  </p>",
      "rawMarkdown": "I am just using a statistical metric (e.g. mean etc.) from all the prediction probabilities. It is one of the biggest challenges  I am facing, as no matter what the metric is, if your case happens (1/300 is FAKE), it is going to cause a big log loss. Again, my model is also not that confident to say, \"yes, I found a FAKE frame, so you call the video FAKE!\". Life is tough!",
      "votes": null
    },
    {
      "id": "774894",
      "postDate": "03/16/2020 03:17:23",
      "content": "<p>Well, maybe you can train your model with biassed loss to make it confident enough to say:</p>\n\n<blockquote>\n  <p>\"yes, I found a FAKE frame, so you call the video FAKE!\"</p>\n</blockquote>",
      "rawMarkdown": "Well, maybe you can train your model with biassed loss to make it confident enough to say:\n&gt; \"yes, I found a FAKE frame, so you call the video FAKE!\"",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 771313,
      "author_name": "debanga",
      "author_url": "",
      "post_date": "03/14/2020 02:22:45",
      "content": "<p>I am just using a statistical metric (e.g. mean etc.) from all the prediction probabilities. It is one of the biggest challenges  I am facing, as no matter what the metric is, if your case happens (1/300 is FAKE), it is going to cause a big log loss. Again, my model is also not that confident to say, \"yes, I found a FAKE frame, so you call the video FAKE!\". Life is tough!  </p>",
      "votes": null,
      "replies": [
        {
          "id": 774894,
          "author_name": "unkownhihi",
          "author_url": "",
          "post_date": "03/16/2020 03:17:23",
          "content": "<p>Well, maybe you can train your model with biassed loss to make it confident enough to say:</p>\n\n<blockquote>\n  <p>\"yes, I found a FAKE frame, so you call the video FAKE!\"</p>\n</blockquote>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "771264": "DDC provided a way to calculate the deep fake result percentage? \nif 1/300 frames are fake Do we determine this is fake or real? \nor just 1/300% as fake?",
    "771313": "I am just using a statistical metric (e.g. mean etc.) from all the prediction probabilities. It is one of the biggest challenges  I am facing, as no matter what the metric is, if your case happens (1/300 is FAKE), it is going to cause a big log loss. Again, my model is also not that confident to say, \"yes, I found a FAKE frame, so you call the video FAKE!\". Life is tough!",
    "774894": "Well, maybe you can train your model with biassed loss to make it confident enough to say:\n&gt; \"yes, I found a FAKE frame, so you call the video FAKE!\""
  },
  "source": "meta"
}