{
  "id": 34182,
  "title": "Best working crop size?",
  "url": "/competitions/intel-mobileodt-cervical-cancer-screening/discussion/34182",
  "author_name": "",
  "post_date": "2017-06-05T11:58:58.107619900Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>What's the best working crop size for you with/without resizing from the center? I've tried using 448x448 crop from the center just to avoid resizing but doesn't work good (in fact in most of the images not a lot of area surrounding cervix is captured). Any magic values which worked well for you? Thanks!</p>",
  "messages": [
    {
      "id": "189166",
      "postDate": "06/05/2017 11:58:58",
      "content": "<p>What's the best working crop size for you with/without resizing from the center? I've tried using 448x448 crop from the center just to avoid resizing but doesn't work good (in fact in most of the images not a lot of area surrounding cervix is captured). Any magic values which worked well for you? Thanks!</p>",
      "rawMarkdown": "What's the best working crop size for you with/without resizing from the center? I've tried using 448x448 crop from the center just to avoid resizing but doesn't work good (in fact in most of the images not a lot of area surrounding cervix is captured). Any magic values which worked well for you? Thanks!",
      "votes": null
    },
    {
      "id": "189219",
      "postDate": "06/05/2017 14:14:45",
      "content": "<p>Just resize. If you resize the entire dataset offline, you don't have to do it online in your data loader / data augmentation. This dataset size is puny (not like the 100GB sealion or 100GB datascience bowl datasets), so just go for it. It's not worth your accuracy score to not do it =) !</p>",
      "rawMarkdown": "Just resize. If you resize the entire dataset offline, you don't have to do it online in your data loader / data augmentation. This dataset size is puny (not like the 100GB sealion or 100GB datascience bowl datasets), so just go for it. It's not worth your accuracy score to not do it =) !",
      "votes": null
    },
    {
      "id": "189254",
      "postDate": "06/05/2017 15:41:56",
      "content": "<p>My first preprocessing step is to crop the original images to square from the centre (eg 4128 x 3096 --&gt; 3096 x 3096).  This seems to capture most of the cervix ROIs, good first cut at reducing the dimensions, without losing too much info.</p>\n\n<p>Second step is to resize to 224 x 224 for feeding into pretrained cnns.  HTH.</p>",
      "rawMarkdown": "My first preprocessing step is to crop the original images to square from the centre (eg 4128 x 3096 --&gt; 3096 x 3096).  This seems to capture most of the cervix ROIs, good first cut at reducing the dimensions, without losing too much info.\n\nSecond step is to resize to 224 x 224 for feeding into pretrained cnns.  HTH.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 189219,
      "author_name": "authman",
      "author_url": "",
      "post_date": "06/05/2017 14:14:45",
      "content": "<p>Just resize. If you resize the entire dataset offline, you don't have to do it online in your data loader / data augmentation. This dataset size is puny (not like the 100GB sealion or 100GB datascience bowl datasets), so just go for it. It's not worth your accuracy score to not do it =) !</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 189254,
      "author_name": "heypeter",
      "author_url": "",
      "post_date": "06/05/2017 15:41:56",
      "content": "<p>My first preprocessing step is to crop the original images to square from the centre (eg 4128 x 3096 --&gt; 3096 x 3096).  This seems to capture most of the cervix ROIs, good first cut at reducing the dimensions, without losing too much info.</p>\n\n<p>Second step is to resize to 224 x 224 for feeding into pretrained cnns.  HTH.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "189166": "What's the best working crop size for you with/without resizing from the center? I've tried using 448x448 crop from the center just to avoid resizing but doesn't work good (in fact in most of the images not a lot of area surrounding cervix is captured). Any magic values which worked well for you? Thanks!",
    "189219": "Just resize. If you resize the entire dataset offline, you don't have to do it online in your data loader / data augmentation. This dataset size is puny (not like the 100GB sealion or 100GB datascience bowl datasets), so just go for it. It's not worth your accuracy score to not do it =) !",
    "189254": "My first preprocessing step is to crop the original images to square from the centre (eg 4128 x 3096 --&gt; 3096 x 3096).  This seems to capture most of the cervix ROIs, good first cut at reducing the dimensions, without losing too much info.\n\nSecond step is to resize to 224 x 224 for feeding into pretrained cnns.  HTH."
  },
  "source": "meta"
}