{
  "id": 18633,
  "title": "Image Preprocessing - Segmentation",
  "url": "/competitions/second-annual-data-science-bowl/discussion/18633",
  "author_name": "",
  "post_date": "2016-01-28T22:38:39.087Z",
  "votes": 1,
  "comment_count": 1,
  "views": 1324,
  "content": "<p>The mxnet and keras tutorials simply crop the image to 64x64. There is no special centering of the heart.</p>\n\n<p>So I wanted to ask if people ranking high on the leaderboard are preprocessing the images (and if how) to center the heart so training a network gets easier?</p>\n\n<p>Any thoughts or findings on <a href=\"https://github.com/Lasagne/Recipes/blob/master/examples/spatial_transformer_network.ipynb\">spatial transformer networks</a>?</p>",
  "messages": [
    {
      "id": "106172",
      "postDate": "01/28/2016 22:38:39",
      "content": "<p>The mxnet and keras tutorials simply crop the image to 64x64. There is no special centering of the heart.</p>\n\n<p>So I wanted to ask if people ranking high on the leaderboard are preprocessing the images (and if how) to center the heart so training a network gets easier?</p>\n\n<p>Any thoughts or findings on <a href=\"https://github.com/Lasagne/Recipes/blob/master/examples/spatial_transformer_network.ipynb\">spatial transformer networks</a>?</p>",
      "rawMarkdown": "The mxnet and keras tutorials simply crop the image to 64x64. There is no special centering of the heart.\r\n\r\nSo I wanted to ask if people ranking high on the leaderboard are preprocessing the images (and if how) to center the heart so training a network gets easier?\r\n\r\nAny thoughts or findings on [spatial transformer networks][1]?\r\n\r\n  [1]: https://github.com/Lasagne/Recipes/blob/master/examples/spatial_transformer_network.ipynb",
      "votes": null
    },
    {
      "id": "106211",
      "postDate": "01/29/2016 04:02:11",
      "content": "<p>@Ren&#233;, For what it's worth, I'm not doing anything nearly that complicated. Something like spatial transforming networks is an interesting idea, but I don't think they would work well without some additional features. Specifically, if the transformer layer is adjusting the image scale, then the second network will lose information about image scale. Since the value we are trying to predict is proportional to square of the image scale, the scale would need to be brought out of the transformer layer somehow and brought back into the second network. I think it's doable, although it does complicate things. I have no idea if it would be worth the trouble though.</p>",
      "rawMarkdown": "@René, For what it's worth, I'm not doing anything nearly that complicated. Something like spatial transforming networks is an interesting idea, but I don't think they would work well without some additional features. Specifically, if the transformer layer is adjusting the image scale, then the second network will lose information about image scale. Since the value we are trying to predict is proportional to square of the image scale, the scale would need to be brought out of the transformer layer somehow and brought back into the second network. I think it's doable, although it does complicate things. I have no idea if it would be worth the trouble though.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 106211,
      "author_name": "bitsofbits",
      "author_url": "",
      "post_date": "01/29/2016 04:02:11",
      "content": "<p>@Ren&#233;, For what it's worth, I'm not doing anything nearly that complicated. Something like spatial transforming networks is an interesting idea, but I don't think they would work well without some additional features. Specifically, if the transformer layer is adjusting the image scale, then the second network will lose information about image scale. Since the value we are trying to predict is proportional to square of the image scale, the scale would need to be brought out of the transformer layer somehow and brought back into the second network. I think it's doable, although it does complicate things. I have no idea if it would be worth the trouble though.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "106172": "The mxnet and keras tutorials simply crop the image to 64x64. There is no special centering of the heart.\r\n\r\nSo I wanted to ask if people ranking high on the leaderboard are preprocessing the images (and if how) to center the heart so training a network gets easier?\r\n\r\nAny thoughts or findings on [spatial transformer networks][1]?\r\n\r\n  [1]: https://github.com/Lasagne/Recipes/blob/master/examples/spatial_transformer_network.ipynb",
    "106211": "@René, For what it's worth, I'm not doing anything nearly that complicated. Something like spatial transforming networks is an interesting idea, but I don't think they would work well without some additional features. Specifically, if the transformer layer is adjusting the image scale, then the second network will lose information about image scale. Since the value we are trying to predict is proportional to square of the image scale, the scale would need to be brought out of the transformer layer somehow and brought back into the second network. I think it's doable, although it does complicate things. I have no idea if it would be worth the trouble though."
  },
  "source": "meta"
}