{
  "id": 207740,
  "title": "Wrong Augmentation",
  "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/207740",
  "author_name": "",
  "post_date": "2020-12-31T04:46:31.228271800Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Why are most of the people doing flip-up/down or left/right augmentation? As it will make the network learn wrong things like the heart in the opposite direction or in the wrong position.</p>",
  "messages": [
    {
      "id": "1133279",
      "postDate": "12/31/2020 04:46:31",
      "content": "<p>Why are most of the people doing flip-up/down or left/right augmentation? As it will make the network learn wrong things like the heart in the opposite direction or in the wrong position.</p>",
      "rawMarkdown": "Why are most of the people doing flip-up/down or left/right augmentation? As it will make the network learn wrong things like the heart in the opposite direction or in the wrong position.",
      "votes": null
    },
    {
      "id": "1133561",
      "postDate": "12/31/2020 10:12:33",
      "content": "<p>If that still teaches the network to correctly recognize whether a catheter is correctly placed, or not, then why not? After all, radiologist might well do alright, if an image is shown upside down, too. In the end, it comes down to trying it and seeing whether it helps as assessed in cross-validation.</p>\n<p>Don't get me wrong, things that don't change the natural \"meaning\" of an image are more obvious first tries.</p>",
      "rawMarkdown": "If that still teaches the network to correctly recognize whether a catheter is correctly placed, or not, then why not? After all, radiologist might well do alright, if an image is shown upside down, too. In the end, it comes down to trying it and seeing whether it helps as assessed in cross-validation.\n\nDon't get me wrong, things that don't change the natural \"meaning\" of an image are more obvious first tries.",
      "votes": null
    },
    {
      "id": "1133738",
      "postDate": "12/31/2020 13:32:42",
      "content": "<p>Augmentation is not only about creating more examples that can be learned by the net, augmentations is another type of <strong>regularization</strong> like <strong>L2/weight decay</strong>, therefore it's final purpose is reduce the likelihood of overfitting the data. Then, even if the images aren't \"realistic\" with flip-up/down (or left/right) augmentation, it's going to work due the reason that i wrote previously.</p>",
      "rawMarkdown": "Augmentation is not only about creating more examples that can be learned by the net, augmentations is another type of **regularization** like **L2/weight decay**, therefore it's final purpose is reduce the likelihood of overfitting the data. Then, even if the images aren't \"realistic\" with flip-up/down (or left/right) augmentation, it's going to work due the reason that i wrote previously.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1133561,
      "author_name": "bjoernholzhauer",
      "author_url": "",
      "post_date": "12/31/2020 10:12:33",
      "content": "<p>If that still teaches the network to correctly recognize whether a catheter is correctly placed, or not, then why not? After all, radiologist might well do alright, if an image is shown upside down, too. In the end, it comes down to trying it and seeing whether it helps as assessed in cross-validation.</p>\n<p>Don't get me wrong, things that don't change the natural \"meaning\" of an image are more obvious first tries.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1133738,
      "author_name": "hiramcho",
      "author_url": "",
      "post_date": "12/31/2020 13:32:42",
      "content": "<p>Augmentation is not only about creating more examples that can be learned by the net, augmentations is another type of <strong>regularization</strong> like <strong>L2/weight decay</strong>, therefore it's final purpose is reduce the likelihood of overfitting the data. Then, even if the images aren't \"realistic\" with flip-up/down (or left/right) augmentation, it's going to work due the reason that i wrote previously.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1133279": "Why are most of the people doing flip-up/down or left/right augmentation? As it will make the network learn wrong things like the heart in the opposite direction or in the wrong position.",
    "1133561": "If that still teaches the network to correctly recognize whether a catheter is correctly placed, or not, then why not? After all, radiologist might well do alright, if an image is shown upside down, too. In the end, it comes down to trying it and seeing whether it helps as assessed in cross-validation.\n\nDon't get me wrong, things that don't change the natural \"meaning\" of an image are more obvious first tries.",
    "1133738": "Augmentation is not only about creating more examples that can be learned by the net, augmentations is another type of **regularization** like **L2/weight decay**, therefore it's final purpose is reduce the likelihood of overfitting the data. Then, even if the images aren't \"realistic\" with flip-up/down (or left/right) augmentation, it's going to work due the reason that i wrote previously."
  },
  "source": "meta"
}