{
  "id": 129109,
  "title": "Denoising frames",
  "url": "/competitions/deepfake-detection-challenge/discussion/129109",
  "author_name": "",
  "post_date": "2020-02-05T14:58:01.437692500Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi, has anyone tested if denoising both real and fake frames causes a significant loss of information?</p>\n\n<p>In my opinion, the ideal scenario is to keep the manipulation artifacts as they are and at the same time removing noise.</p>\n\n<p>Denoising images using\n<code>cv2.fastNlMeansDenoising(img, h=5.0, templateWindowSize=7, searchWindowSize=21)</code></p>\n\n<p>I get </p>\n\n<p>Original:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F13e7d38ea536ec403e5c23e80e0af687%2Ftemp.png?generation=1580914048598928&amp;alt=media\" alt=\"\"></p>\n\n<p>Denoised:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F98e1c859e13c2b42b795e31d2ba94572%2Ftemp1.png?generation=1580914048055206&amp;alt=media\" alt=\"\"></p>\n\n<p>and for a real one</p>\n\n<p>Original:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2Fe6478ae8d5dfee87e93bc584c8a594e8%2Ftemp3.png?generation=1580914105101835&amp;alt=media\" alt=\"\"></p>\n\n<p>Denoised:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F46b951cf826bcd415014b8cd4934676c%2Ftemp2.png?generation=1580914105400809&amp;alt=media\" alt=\"\"></p>\n\n<p>Maybe setting a smaller h might be better.</p>",
  "messages": [
    {
      "id": "737591",
      "postDate": "02/05/2020 14:58:01",
      "content": "<p>Hi, has anyone tested if denoising both real and fake frames causes a significant loss of information?</p>\n\n<p>In my opinion, the ideal scenario is to keep the manipulation artifacts as they are and at the same time removing noise.</p>\n\n<p>Denoising images using\n<code>cv2.fastNlMeansDenoising(img, h=5.0, templateWindowSize=7, searchWindowSize=21)</code></p>\n\n<p>I get </p>\n\n<p>Original:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F13e7d38ea536ec403e5c23e80e0af687%2Ftemp.png?generation=1580914048598928&amp;alt=media\" alt=\"\"></p>\n\n<p>Denoised:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F98e1c859e13c2b42b795e31d2ba94572%2Ftemp1.png?generation=1580914048055206&amp;alt=media\" alt=\"\"></p>\n\n<p>and for a real one</p>\n\n<p>Original:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2Fe6478ae8d5dfee87e93bc584c8a594e8%2Ftemp3.png?generation=1580914105101835&amp;alt=media\" alt=\"\"></p>\n\n<p>Denoised:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F46b951cf826bcd415014b8cd4934676c%2Ftemp2.png?generation=1580914105400809&amp;alt=media\" alt=\"\"></p>\n\n<p>Maybe setting a smaller h might be better.</p>",
      "rawMarkdown": "Hi, has anyone tested if denoising both real and fake frames causes a significant loss of information?\n\nIn my opinion, the ideal scenario is to keep the manipulation artifacts as they are and at the same time removing noise.\n\nDenoising images using\n`cv2.fastNlMeansDenoising(img, h=5.0, templateWindowSize=7, searchWindowSize=21)`\n\nI get \n\nOriginal:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F13e7d38ea536ec403e5c23e80e0af687%2Ftemp.png?generation=1580914048598928&amp;alt=media)\n\nDenoised:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F98e1c859e13c2b42b795e31d2ba94572%2Ftemp1.png?generation=1580914048055206&amp;alt=media)\n\nand for a real one\n\nOriginal:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2Fe6478ae8d5dfee87e93bc584c8a594e8%2Ftemp3.png?generation=1580914105101835&amp;alt=media)\n\nDenoised:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F46b951cf826bcd415014b8cd4934676c%2Ftemp2.png?generation=1580914105400809&amp;alt=media)\n\nMaybe setting a smaller h might be better.",
      "votes": null
    },
    {
      "id": "737808",
      "postDate": "02/05/2020 19:45:44",
      "content": "<p>Interesting. This probably won't work very well with good fakes but there are a lot of lousy fakes in this dataset, and you might be able to easily detect them using this method.</p>",
      "rawMarkdown": "Interesting. This probably won't work very well with good fakes but there are a lot of lousy fakes in this dataset, and you might be able to easily detect them using this method.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 737808,
      "author_name": "humananalog",
      "author_url": "",
      "post_date": "02/05/2020 19:45:44",
      "content": "<p>Interesting. This probably won't work very well with good fakes but there are a lot of lousy fakes in this dataset, and you might be able to easily detect them using this method.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "737591": "Hi, has anyone tested if denoising both real and fake frames causes a significant loss of information?\n\nIn my opinion, the ideal scenario is to keep the manipulation artifacts as they are and at the same time removing noise.\n\nDenoising images using\n`cv2.fastNlMeansDenoising(img, h=5.0, templateWindowSize=7, searchWindowSize=21)`\n\nI get \n\nOriginal:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F13e7d38ea536ec403e5c23e80e0af687%2Ftemp.png?generation=1580914048598928&amp;alt=media)\n\nDenoised:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F98e1c859e13c2b42b795e31d2ba94572%2Ftemp1.png?generation=1580914048055206&amp;alt=media)\n\nand for a real one\n\nOriginal:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2Fe6478ae8d5dfee87e93bc584c8a594e8%2Ftemp3.png?generation=1580914105101835&amp;alt=media)\n\nDenoised:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3739757%2F46b951cf826bcd415014b8cd4934676c%2Ftemp2.png?generation=1580914105400809&amp;alt=media)\n\nMaybe setting a smaller h might be better.",
    "737808": "Interesting. This probably won't work very well with good fakes but there are a lot of lousy fakes in this dataset, and you might be able to easily detect them using this method."
  },
  "source": "meta"
}