{
  "id": 290303,
  "title": "What pre-processing have you found to be effective?",
  "url": "/competitions/applications-of-deep-learning-wustlfall-2021/discussion/290303",
  "author_name": "Fa C Shus",
  "post_date": "2021-11-24T01:27:33.467000",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have been experimenting with the ImageDataGenerator properties:<br>\nHorizontal flip<br>\nrotation <br>\ngrayscale</p>\n<p>I'm so new to image processing that I cannot yet:</p>\n<ul>\n<li>effectively drop out the glare from the light</li>\n<li>remove the background sky nor ground</li>\n</ul>",
  "messages": [
    {
      "id": 1593476,
      "postDate": "2021-11-24T01:27:33.467Z",
      "content": "<p>I have been experimenting with the ImageDataGenerator properties:<br>\nHorizontal flip<br>\nrotation <br>\ngrayscale</p>\n<p>I'm so new to image processing that I cannot yet:</p>\n<ul>\n<li>effectively drop out the glare from the light</li>\n<li>remove the background sky nor ground</li>\n</ul>",
      "rawMarkdown": "I have been experimenting with the ImageDataGenerator properties:\nHorizontal flip\nrotation \ngrayscale\n\nI'm so new to image processing that I cannot yet:\n- effectively drop out the glare from the light\n- remove the background sky nor ground",
      "votes": 1
    },
    {
      "id": 1594766,
      "postDate": "2021-11-25T05:46:46.927Z",
      "content": "<p>Hi Fa,</p>\n<p>I've had the following thoughts on data augmentation (pre-processing):</p>\n<p>Horizontal flip: This should be effective as it doubles the training data.<br>\nGrayscale: As the building colors are not important, this should be OK to do.<br>\nRotation: Have not tried, I'm guessing this would not be effective as all the buildings are vertical, and none of the images in the scoring dataset will be rotated.</p>\n<p>I have looked at zooming in and out as a transform, some of the smaller images could be zoomed to create more training data.</p>\n<p>I'd also be keen to hear more thoughts on this…</p>\n<p>Regards</p>\n<p>Alan</p>",
      "rawMarkdown": "Hi Fa,\n\nI've had the following thoughts on data augmentation (pre-processing):\n\nHorizontal flip: This should be effective as it doubles the training data.\nGrayscale: As the building colors are not important, this should be OK to do.\nRotation: Have not tried, I'm guessing this would not be effective as all the buildings are vertical, and none of the images in the scoring dataset will be rotated.\n\nI have looked at zooming in and out as a transform, some of the smaller images could be zoomed to create more training data.\n\nI'd also be keen to hear more thoughts on this...\n\nRegards\n\nAlan\n\n\n"
    },
    {
      "id": 1594438,
      "postDate": "2021-11-24T20:36:50.883Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1594766,
      "author_name": "Alan Smith",
      "author_url": "",
      "post_date": "2021-11-25T05:46:46.927000",
      "content": "<p>Hi Fa,</p>\n<p>I've had the following thoughts on data augmentation (pre-processing):</p>\n<p>Horizontal flip: This should be effective as it doubles the training data.<br>\nGrayscale: As the building colors are not important, this should be OK to do.<br>\nRotation: Have not tried, I'm guessing this would not be effective as all the buildings are vertical, and none of the images in the scoring dataset will be rotated.</p>\n<p>I have looked at zooming in and out as a transform, some of the smaller images could be zoomed to create more training data.</p>\n<p>I'd also be keen to hear more thoughts on this…</p>\n<p>Regards</p>\n<p>Alan</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1594438,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-11-24T20:36:50.883000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1593476": "I have been experimenting with the ImageDataGenerator properties:\nHorizontal flip\nrotation \ngrayscale\n\nI'm so new to image processing that I cannot yet:\n- effectively drop out the glare from the light\n- remove the background sky nor ground",
    "1594766": "Hi Fa,\n\nI've had the following thoughts on data augmentation (pre-processing):\n\nHorizontal flip: This should be effective as it doubles the training data.\nGrayscale: As the building colors are not important, this should be OK to do.\nRotation: Have not tried, I'm guessing this would not be effective as all the buildings are vertical, and none of the images in the scoring dataset will be rotated.\n\nI have looked at zooming in and out as a transform, some of the smaller images could be zoomed to create more training data.\n\nI'd also be keen to hear more thoughts on this...\n\nRegards\n\nAlan\n\n\n",
    "1594438": ""
  }
}