{
  "id": 233834,
  "title": "Width/height ratios of enc input res around 2 most 'efficient'",
  "url": "/competitions/bms-molecular-translation/discussion/233834",
  "author_name": "Michael Wolff",
  "post_date": "2021-04-21T11:27:10.837000",
  "votes": 12,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi guys,<br>\nit's my first kaggle competition and I wanted to contribute after learning so much already from the shared kernels!<br>\nI've analysed the image sizes after cropping to find 'efficient' encoder input resolutions: Width / height ratios around 2 fit the images most efficient (I rotated all images to be horizontal). I got a slight LS distance improvement of 0.4 even from switching from 448x256 to 512x256.</p>\n<p>I put together a table with the following statistics for different typical resolutions and width/height ratios:</p>\n<ul>\n<li>the fraction of cropped images that need to be resized to fit the given resolution ('shrinked')</li>\n<li>the mean shrink factor</li>\n<li>the mean shrink factor of only those images that need to be shrinked</li>\n<li>and some more</li>\n</ul>\n<p>You can find the results at the end of <a href=\"https://www.kaggle.com/michaelwolff/bms-inchi-cropped-img-sizes-for-best-resolution\" target=\"_blank\">this kernel</a>.</p>",
  "messages": [
    {
      "id": 1279921,
      "postDate": "2021-04-21T11:27:10.837Z",
      "content": "<p>Hi guys,<br>\nit's my first kaggle competition and I wanted to contribute after learning so much already from the shared kernels!<br>\nI've analysed the image sizes after cropping to find 'efficient' encoder input resolutions: Width / height ratios around 2 fit the images most efficient (I rotated all images to be horizontal). I got a slight LS distance improvement of 0.4 even from switching from 448x256 to 512x256.</p>\n<p>I put together a table with the following statistics for different typical resolutions and width/height ratios:</p>\n<ul>\n<li>the fraction of cropped images that need to be resized to fit the given resolution ('shrinked')</li>\n<li>the mean shrink factor</li>\n<li>the mean shrink factor of only those images that need to be shrinked</li>\n<li>and some more</li>\n</ul>\n<p>You can find the results at the end of <a href=\"https://www.kaggle.com/michaelwolff/bms-inchi-cropped-img-sizes-for-best-resolution\" target=\"_blank\">this kernel</a>.</p>",
      "rawMarkdown": "Hi guys,\nit's my first kaggle competition and I wanted to contribute after learning so much already from the shared kernels!\nI've analysed the image sizes after cropping to find 'efficient' encoder input resolutions: Width / height ratios around 2 fit the images most efficient (I rotated all images to be horizontal). I got a slight LS distance improvement of 0.4 even from switching from 448x256 to 512x256.\n\nI put together a table with the following statistics for different typical resolutions and width/height ratios:\n- the fraction of cropped images that need to be resized to fit the given resolution ('shrinked')\n- the mean shrink factor\n- the mean shrink factor of only those images that need to be shrinked\n- and some more\n\nYou can find the results at the end of [this kernel](https://www.kaggle.com/michaelwolff/bms-inchi-cropped-img-sizes-for-best-resolution).\n",
      "votes": 12
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1279921": "Hi guys,\nit's my first kaggle competition and I wanted to contribute after learning so much already from the shared kernels!\nI've analysed the image sizes after cropping to find 'efficient' encoder input resolutions: Width / height ratios around 2 fit the images most efficient (I rotated all images to be horizontal). I got a slight LS distance improvement of 0.4 even from switching from 448x256 to 512x256.\n\nI put together a table with the following statistics for different typical resolutions and width/height ratios:\n- the fraction of cropped images that need to be resized to fit the given resolution ('shrinked')\n- the mean shrink factor\n- the mean shrink factor of only those images that need to be shrinked\n- and some more\n\nYou can find the results at the end of [this kernel](https://www.kaggle.com/michaelwolff/bms-inchi-cropped-img-sizes-for-best-resolution).\n"
  }
}