{
  "id": 167015,
  "title": "Cropping and Hue augmentations. Are they really useful?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/167015",
  "author_name": "",
  "post_date": "2020-07-14T21:43:41.340093400Z",
  "votes": null,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I'm specially afraid of these two augmentations. Since not all of the images are perfectly centered I don't know if cropping would be useful.</p>\n\n<p>And what are your opinions on Hue? My intuition says that it could modify the images to the point that it give us \"mislabeled\" data. But I'm just assuming that.</p>\n\n<p>Does anyone has interesting takes on this?</p>",
  "messages": [
    {
      "id": "929702",
      "postDate": "07/14/2020 21:43:41",
      "content": "<p>I'm specially afraid of these two augmentations. Since not all of the images are perfectly centered I don't know if cropping would be useful.</p>\n\n<p>And what are your opinions on Hue? My intuition says that it could modify the images to the point that it give us \"mislabeled\" data. But I'm just assuming that.</p>\n\n<p>Does anyone has interesting takes on this?</p>",
      "rawMarkdown": "I'm specially afraid of these two augmentations. Since not all of the images are perfectly centered I don't know if cropping would be useful.\n\nAnd what are your opinions on Hue? My intuition says that it could modify the images to the point that it give us \"mislabeled\" data. But I'm just assuming that.\n\nDoes anyone has interesting takes on this?",
      "votes": null
    },
    {
      "id": "929704",
      "postDate": "07/14/2020 21:46:34",
      "content": "<p>There is only one way to find out -- try it!</p>",
      "rawMarkdown": "There is only one way to find out -- try it!",
      "votes": null
    },
    {
      "id": "929718",
      "postDate": "07/14/2020 22:27:14",
      "content": "<p>Good answer haha! Indeed. From my previous experiments I got better LB scores without cropping and Hue augmentations. But I was really curious because there are a lot of public kernels with these augmentations so maybe I'm not tuning them correctly. </p>",
      "rawMarkdown": "Good answer haha! Indeed. From my previous experiments I got better LB scores without cropping and Hue augmentations. But I was really curious because there are a lot of public kernels with these augmentations so maybe I'm not tuning them correctly.",
      "votes": null
    },
    {
      "id": "929848",
      "postDate": "07/15/2020 02:47:25",
      "content": "<p>Same I got good LB when not using random cropping and random hue. I too have the same reason for not using them as you have. \nAlso, I'm not using shear augmentation I think it might produce images with asymmetric boundaries in begin cases.</p>",
      "rawMarkdown": "Same I got good LB when not using random cropping and random hue. I too have the same reason for not using them as you have. \nAlso, I'm not using shear augmentation I think it might produce images with asymmetric boundaries in begin cases.",
      "votes": null
    },
    {
      "id": "929854",
      "postDate": "07/15/2020 02:57:53",
      "content": "<p>Hum makes sense! Gonna try without my current shears to see how it goes.</p>",
      "rawMarkdown": "Hum makes sense! Gonna try without my current shears to see how it goes.",
      "votes": null
    },
    {
      "id": "930084",
      "postDate": "07/15/2020 07:20:03",
      "content": "<p>how to do random cropping in tensorflow while reading through tfrec??</p>",
      "rawMarkdown": "how to do random cropping in tensorflow while reading through tfrec??",
      "votes": null
    },
    {
      "id": "930147",
      "postDate": "07/15/2020 08:23:18",
      "content": "<p><code>\ntf.image.random_crop(\n    image, [crop_height, crop_width, 3] , seed=None\n)\n</code></p>",
      "rawMarkdown": "```\ntf.image.random_crop(\n    image, [crop_height, crop_width, 3] , seed=None\n)\n```",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 929704,
      "author_name": "graf10a",
      "author_url": "",
      "post_date": "07/14/2020 21:46:34",
      "content": "<p>There is only one way to find out -- try it!</p>",
      "votes": null,
      "replies": [
        {
          "id": 929718,
          "author_name": "santiviquez",
          "author_url": "",
          "post_date": "07/14/2020 22:27:14",
          "content": "<p>Good answer haha! Indeed. From my previous experiments I got better LB scores without cropping and Hue augmentations. But I was really curious because there are a lot of public kernels with these augmentations so maybe I'm not tuning them correctly. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 929848,
          "author_name": "prateek0x",
          "author_url": "",
          "post_date": "07/15/2020 02:47:25",
          "content": "<p>Same I got good LB when not using random cropping and random hue. I too have the same reason for not using them as you have. \nAlso, I'm not using shear augmentation I think it might produce images with asymmetric boundaries in begin cases.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 929854,
          "author_name": "santiviquez",
          "author_url": "",
          "post_date": "07/15/2020 02:57:53",
          "content": "<p>Hum makes sense! Gonna try without my current shears to see how it goes.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 930084,
      "author_name": "msharuk589",
      "author_url": "",
      "post_date": "07/15/2020 07:20:03",
      "content": "<p>how to do random cropping in tensorflow while reading through tfrec??</p>",
      "votes": null,
      "replies": [
        {
          "id": 930147,
          "author_name": "prateek0x",
          "author_url": "",
          "post_date": "07/15/2020 08:23:18",
          "content": "<p><code>\ntf.image.random_crop(\n    image, [crop_height, crop_width, 3] , seed=None\n)\n</code></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "929702": "I'm specially afraid of these two augmentations. Since not all of the images are perfectly centered I don't know if cropping would be useful.\n\nAnd what are your opinions on Hue? My intuition says that it could modify the images to the point that it give us \"mislabeled\" data. But I'm just assuming that.\n\nDoes anyone has interesting takes on this?",
    "929704": "There is only one way to find out -- try it!",
    "929718": "Good answer haha! Indeed. From my previous experiments I got better LB scores without cropping and Hue augmentations. But I was really curious because there are a lot of public kernels with these augmentations so maybe I'm not tuning them correctly.",
    "929848": "Same I got good LB when not using random cropping and random hue. I too have the same reason for not using them as you have. \nAlso, I'm not using shear augmentation I think it might produce images with asymmetric boundaries in begin cases.",
    "929854": "Hum makes sense! Gonna try without my current shears to see how it goes.",
    "930084": "how to do random cropping in tensorflow while reading through tfrec??",
    "930147": "```\ntf.image.random_crop(\n    image, [crop_height, crop_width, 3] , seed=None\n)\n```"
  },
  "source": "meta"
}