{
  "id": 12817,
  "title": "What image sizes are you using?",
  "url": "/competitions/diabetic-retinopathy-detection/discussion/12817",
  "author_name": "",
  "post_date": "2015-03-15T14:27:27.537Z",
  "votes": 4,
  "comment_count": 3,
  "views": 2076,
  "content": "<p>It's pretty much a given that you have to down-scale the images in this competition (they are huge!) - but to what size?</p>\n<p>Smaller images of course means faster training (less computation, memory usage etc), but of course also loss of quality. Just from manual inspection I feel than anything smaller than 100x100 starts losing too much information to be useful, though I could be wrong. Does anyone mind sharing their thoughts on this issue?</p>",
  "messages": [
    {
      "id": "66235",
      "postDate": "03/15/2015 14:27:27",
      "content": "<p>It's pretty much a given that you have to down-scale the images in this competition (they are huge!) - but to what size?</p>\n<p>Smaller images of course means faster training (less computation, memory usage etc), but of course also loss of quality. Just from manual inspection I feel than anything smaller than 100x100 starts losing too much information to be useful, though I could be wrong. Does anyone mind sharing their thoughts on this issue?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "66293",
      "postDate": "03/15/2015 22:09:39",
      "content": "<p>Right now I'm testing the convolutional neural networks which is fed with 425 x 425 patches cropped from images resized to 512 x 512 px.</p>\n<p>Before that I tried with 1024 x 1024 but the results weren't any better.</p>\n<p>Unfortunately the accuracy of the current model doesn't improve after 1000 iterations and am trying to figure out how the resolution affects the performance.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "67500",
      "postDate": "03/21/2015 05:45:08",
      "content": "<p>Maybe it's worth considering the size of the smallest feature. At stage 1 or 2, there are microaneurysms visible, but they are really just little dots of a few pixels, even in the original images. Perhaps there is a clever way to identify plainly uninteresting segments of the image before feeding the more interesting segments to a convnet....</p>\n\n<p>There is a description of the features corresponding to each scoring level in this paper (thanks to Deep for finding it)</p>\n<p><a href=\"https://storage.googleapis.com/kaggle-forum-message-attachments/66827/2169/Algorithms for the Automated Detection of Diabetic Retinopathy Using Digital Fundus Images - A Review.pdf?sv=2012-02-12&se=2015-03-24T01:55:11Z&sr=b&sp=r&sig=5tm9dDHSTxY4XVffPDyxwI37%2FBxQUTE20U1%2FaYCEc54%3D\" target=\"_blank\">Algorithms for the Automated Detection of Diabetic Retinopathy Using Digital Fundus Images: A Review</a></p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "67534",
      "postDate": "03/21/2015 16:24:06",
      "content": "<p>Yeah, just dumping these images into a convnet is probably not going&nbsp;to be the most effective approach. &nbsp;There is a lot of structure to take advantage of.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 66293,
      "author_name": "nikogamulin",
      "author_url": "",
      "post_date": "03/15/2015 22:09:39",
      "content": "<p>Right now I'm testing the convolutional neural networks which is fed with 425 x 425 patches cropped from images resized to 512 x 512 px.</p>\n<p>Before that I tried with 1024 x 1024 but the results weren't any better.</p>\n<p>Unfortunately the accuracy of the current model doesn't improve after 1000 iterations and am trying to figure out how the resolution affects the performance.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 67500,
      "author_name": "smallyellowduck",
      "author_url": "",
      "post_date": "03/21/2015 05:45:08",
      "content": "<p>Maybe it's worth considering the size of the smallest feature. At stage 1 or 2, there are microaneurysms visible, but they are really just little dots of a few pixels, even in the original images. Perhaps there is a clever way to identify plainly uninteresting segments of the image before feeding the more interesting segments to a convnet....</p>\n\n<p>There is a description of the features corresponding to each scoring level in this paper (thanks to Deep for finding it)</p>\n<p><a href=\"https://storage.googleapis.com/kaggle-forum-message-attachments/66827/2169/Algorithms for the Automated Detection of Diabetic Retinopathy Using Digital Fundus Images - A Review.pdf?sv=2012-02-12&se=2015-03-24T01:55:11Z&sr=b&sp=r&sig=5tm9dDHSTxY4XVffPDyxwI37%2FBxQUTE20U1%2FaYCEc54%3D\" target=\"_blank\">Algorithms for the Automated Detection of Diabetic Retinopathy Using Digital Fundus Images: A Review</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 67534,
      "author_name": "alexcoventry",
      "author_url": "",
      "post_date": "03/21/2015 16:24:06",
      "content": "<p>Yeah, just dumping these images into a convnet is probably not going&nbsp;to be the most effective approach. &nbsp;There is a lot of structure to take advantage of.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "66235": "",
    "66293": "",
    "67500": "",
    "67534": ""
  },
  "source": "meta"
}