{
  "id": 140246,
  "title": "Time to unveil esoteric methods :)",
  "url": "/competitions/deepfake-detection-challenge/discussion/140246",
  "author_name": "",
  "post_date": "2020-04-01T02:56:43.568042500Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>It's the end of submission deadline.\nAnd I'm so excited to expect to some of the top ranker would share their expeditious key methods - plz :)</p>\n\n<p>And start with my humble trial.</p>\n\n<p>I marked 0.34250.\nAnd I'll list my own method in order of priority on my thoughts.</p>\n\n<p>1) Making up dataset\n   I lately read in discussion board that most effective way to make training dataset is just pick out 1-frame from 1-video to avoid overfitting to specific human face.\n  But I pick 5-frame from all fake video, and 25-frame for all real video to balance the number.\n  So I prepared 920,000 frames for training ( keep 280,000 frames of 10 folder for validation)</p>\n\n<p>2) Pre-processing\n  I started with minimal pre-processing including cropping, vertical flips.\n  And add color variation, random jpeg quality, shear transformation.</p>\n\n<p>3) Model\n   - My first model is modified Deeplap V3 plus for multi-task learning \n     ( Loss of Semantic segmantation &amp; Classification )\n     cuz I thought each pixel-difference original from manipulated would help to boost training speed.\n   It works in early stage of training, but I think it was not so helpful in the late stage for more than 2~3epochs, because I couldn't get accurate semantic label for each pixel due to the disturbance noise of video images.</p>\n\n<pre><code>- And I add efficientnet-b2 / b3 for ensemble method.\n</code></pre>\n\n<p>&gt; I guess key ideas for improvement might be in 1) making dataset 2) Pre-processing</p>\n\n<p>I really expect any participants will discuss about their idea in this board.</p>",
  "messages": [
    {
      "id": "793431",
      "postDate": "04/01/2020 02:56:43",
      "content": "<p>It's the end of submission deadline.\nAnd I'm so excited to expect to some of the top ranker would share their expeditious key methods - plz :)</p>\n\n<p>And start with my humble trial.</p>\n\n<p>I marked 0.34250.\nAnd I'll list my own method in order of priority on my thoughts.</p>\n\n<p>1) Making up dataset\n   I lately read in discussion board that most effective way to make training dataset is just pick out 1-frame from 1-video to avoid overfitting to specific human face.\n  But I pick 5-frame from all fake video, and 25-frame for all real video to balance the number.\n  So I prepared 920,000 frames for training ( keep 280,000 frames of 10 folder for validation)</p>\n\n<p>2) Pre-processing\n  I started with minimal pre-processing including cropping, vertical flips.\n  And add color variation, random jpeg quality, shear transformation.</p>\n\n<p>3) Model\n   - My first model is modified Deeplap V3 plus for multi-task learning \n     ( Loss of Semantic segmantation &amp; Classification )\n     cuz I thought each pixel-difference original from manipulated would help to boost training speed.\n   It works in early stage of training, but I think it was not so helpful in the late stage for more than 2~3epochs, because I couldn't get accurate semantic label for each pixel due to the disturbance noise of video images.</p>\n\n<pre><code>- And I add efficientnet-b2 / b3 for ensemble method.\n</code></pre>\n\n<p>&gt; I guess key ideas for improvement might be in 1) making dataset 2) Pre-processing</p>\n\n<p>I really expect any participants will discuss about their idea in this board.</p>",
      "rawMarkdown": "It's the end of submission deadline.\nAnd I'm so excited to expect to some of the top ranker would share their expeditious key methods - plz :)\n\nAnd start with my humble trial.\n\nI marked 0.34250.\nAnd I'll list my own method in order of priority on my thoughts.\n\n1) Making up dataset\n   I lately read in discussion board that most effective way to make training dataset is just pick out 1-frame from 1-video to avoid overfitting to specific human face.\n  But I pick 5-frame from all fake video, and 25-frame for all real video to balance the number.\n  So I prepared 920,000 frames for training ( keep 280,000 frames of 10 folder for validation)\n\n2) Pre-processing\n  I started with minimal pre-processing including cropping, vertical flips.\n  And add color variation, random jpeg quality, shear transformation.\n\n 3) Model\n   - My first model is modified Deeplap V3 plus for multi-task learning \n     ( Loss of Semantic segmantation &amp; Classification )\n     cuz I thought each pixel-difference original from manipulated would help to boost training speed.\n   It works in early stage of training, but I think it was not so helpful in the late stage for more than 2~3epochs, because I couldn't get accurate semantic label for each pixel due to the disturbance noise of video images.\n   \n    - And I add efficientnet-b2 / b3 for ensemble method.\n\n&gt; I guess key ideas for improvement might be in 1) making dataset 2) Pre-processing\n\nI really expect any participants will discuss about their idea in this board.",
      "votes": null
    },
    {
      "id": "793438",
      "postDate": "04/01/2020 03:05:43",
      "content": "<p>Using more data is equivalent to data augmentation. I think you meant 25 frames from real videos and 5 from fake videos.</p>",
      "rawMarkdown": "Using more data is equivalent to data augmentation. I think you meant 25 frames from real videos and 5 from fake videos.",
      "votes": null
    },
    {
      "id": "793478",
      "postDate": "04/01/2020 04:00:36",
      "content": "<p>You're right. I'll modify the content as you pointed :)</p>",
      "rawMarkdown": "You're right. I'll modify the content as you pointed :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 793438,
      "author_name": "yuanzhezhou",
      "author_url": "",
      "post_date": "04/01/2020 03:05:43",
      "content": "<p>Using more data is equivalent to data augmentation. I think you meant 25 frames from real videos and 5 from fake videos.</p>",
      "votes": null,
      "replies": [
        {
          "id": 793478,
          "author_name": "gwsong",
          "author_url": "",
          "post_date": "04/01/2020 04:00:36",
          "content": "<p>You're right. I'll modify the content as you pointed :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "793431": "It's the end of submission deadline.\nAnd I'm so excited to expect to some of the top ranker would share their expeditious key methods - plz :)\n\nAnd start with my humble trial.\n\nI marked 0.34250.\nAnd I'll list my own method in order of priority on my thoughts.\n\n1) Making up dataset\n   I lately read in discussion board that most effective way to make training dataset is just pick out 1-frame from 1-video to avoid overfitting to specific human face.\n  But I pick 5-frame from all fake video, and 25-frame for all real video to balance the number.\n  So I prepared 920,000 frames for training ( keep 280,000 frames of 10 folder for validation)\n\n2) Pre-processing\n  I started with minimal pre-processing including cropping, vertical flips.\n  And add color variation, random jpeg quality, shear transformation.\n\n 3) Model\n   - My first model is modified Deeplap V3 plus for multi-task learning \n     ( Loss of Semantic segmantation &amp; Classification )\n     cuz I thought each pixel-difference original from manipulated would help to boost training speed.\n   It works in early stage of training, but I think it was not so helpful in the late stage for more than 2~3epochs, because I couldn't get accurate semantic label for each pixel due to the disturbance noise of video images.\n   \n    - And I add efficientnet-b2 / b3 for ensemble method.\n\n&gt; I guess key ideas for improvement might be in 1) making dataset 2) Pre-processing\n\nI really expect any participants will discuss about their idea in this board.",
    "793438": "Using more data is equivalent to data augmentation. I think you meant 25 frames from real videos and 5 from fake videos.",
    "793478": "You're right. I'll modify the content as you pointed :)"
  },
  "source": "meta"
}