{
  "id": 26508,
  "title": "Manual steps allowed?",
  "url": "/competitions/dstl-satellite-imagery-feature-detection/discussion/26508",
  "author_name": "",
  "post_date": "2016-12-15T01:57:28.583Z",
  "votes": 18,
  "comment_count": 1,
  "views": 1253,
  "content": "<p>I'm just wondering if manuals step are allowed? (they could be programmed, but it takes a while to code them).</p>\n\n<p>It doesn't seem to violate the rules:</p>\n\n<blockquote>\n  <p>Submissions may not use or incorporate information from hand labeling\n  or human prediction of the validation dataset or test data records.</p>\n</blockquote>\n\n<p>But is not receivable in case of a winning solution, right?:</p>\n\n<blockquote>\n  <p>The delivered software code must be capable of generating the winning\n  Submission and contain a description of resources required to build\n  and/or run the executable code successfully.</p>\n</blockquote>\n\n<p>For instance taking the base picture, I want the roads:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>I can apply a Kuwahara filter, which makes the roads much easier to discern.</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5583/kuwahara_filter.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Then I transform to 32-bit colors, and perform unsupervised connection region computation:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5575/connected_regions.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Picking back the 32-bit color picture, I convert back to 8-bit, and perform masking by nearby points algorithms:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5580/within_distance_masking.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Turned to a skeleton:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5578/skeleton_mask.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>I can also use Suliman's fuzzy edge detection filter:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5579/suliman_edge.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Bringing back visual using gradient direction:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5581/gradient_skeleton.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Roads are much easier to notice here with a skeleton:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5576/gradient_direction.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Just a synthesis picture after inverting the skeleton, masking the kuwahara+connected regions, and stacking these 2 layers with the original pic:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5577/final.jpg\" alt=\"enter image description here\" title=\"\"><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>The level of details is low here because the picture is small, but on a large picture it is more noticeable. If I'm not mistaken, nearly all roads were covered, except some parts of large roads and the very thin ones which requires parameter tuning of filters / algorithms used.</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5585/stack.gif\" alt=\"enter image description here\" title=\"\"></p>",
  "messages": [
    {
      "id": "150403",
      "postDate": "12/15/2016 01:57:28",
      "content": "<p>I'm just wondering if manuals step are allowed? (they could be programmed, but it takes a while to code them).</p>\n\n<p>It doesn't seem to violate the rules:</p>\n\n<blockquote>\n  <p>Submissions may not use or incorporate information from hand labeling\n  or human prediction of the validation dataset or test data records.</p>\n</blockquote>\n\n<p>But is not receivable in case of a winning solution, right?:</p>\n\n<blockquote>\n  <p>The delivered software code must be capable of generating the winning\n  Submission and contain a description of resources required to build\n  and/or run the executable code successfully.</p>\n</blockquote>\n\n<p>For instance taking the base picture, I want the roads:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>I can apply a Kuwahara filter, which makes the roads much easier to discern.</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5583/kuwahara_filter.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Then I transform to 32-bit colors, and perform unsupervised connection region computation:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5575/connected_regions.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Picking back the 32-bit color picture, I convert back to 8-bit, and perform masking by nearby points algorithms:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5580/within_distance_masking.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Turned to a skeleton:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5578/skeleton_mask.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>I can also use Suliman's fuzzy edge detection filter:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5579/suliman_edge.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Bringing back visual using gradient direction:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5581/gradient_skeleton.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Roads are much easier to notice here with a skeleton:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5576/gradient_direction.jpg\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>Just a synthesis picture after inverting the skeleton, masking the kuwahara+connected regions, and stacking these 2 layers with the original pic:</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5577/final.jpg\" alt=\"enter image description here\" title=\"\"><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>The level of details is low here because the picture is small, but on a large picture it is more noticeable. If I'm not mistaken, nearly all roads were covered, except some parts of large roads and the very thin ones which requires parameter tuning of filters / algorithms used.</p>\n\n<p><img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5585/stack.gif\" alt=\"enter image description here\" title=\"\"></p>",
      "rawMarkdown": "I'm just wondering if manuals step are allowed? (they could be programmed, but it takes a while to code them).\r\n\r\nIt doesn't seem to violate the rules:\r\n\r\n> Submissions may not use or incorporate information from hand labeling\r\n> or human prediction of the validation dataset or test data records.\r\n\r\nBut is not receivable in case of a winning solution, right?:\r\n\r\n> The delivered software code must be capable of generating the winning\r\n> Submission and contain a description of resources required to build\r\n> and/or run the executable code successfully.\r\n\r\nFor instance taking the base picture, I want the roads:\r\n\r\n![enter image description here][1]\r\n\r\nI can apply a Kuwahara filter, which makes the roads much easier to discern.\r\n\r\n![enter image description here][2]\r\n\r\nThen I transform to 32-bit colors, and perform unsupervised connection region computation:\r\n\r\n![enter image description here][3]\r\n\r\nPicking back the 32-bit color picture, I convert back to 8-bit, and perform masking by nearby points algorithms:\r\n\r\n![enter image description here][4]\r\n\r\nTurned to a skeleton:\r\n\r\n![enter image description here][5]\r\n\r\nI can also use Suliman's fuzzy edge detection filter:\r\n\r\n![enter image description here][6]\r\n\r\nBringing back visual using gradient direction:\r\n\r\n![enter image description here][7]\r\n\r\nRoads are much easier to notice here with a skeleton:\r\n\r\n![enter image description here][8]\r\n\r\nJust a synthesis picture after inverting the skeleton, masking the kuwahara+connected regions, and stacking these 2 layers with the original pic:\r\n\r\n![enter image description here][9]![enter image description here][10]\r\n\r\nThe level of details is low here because the picture is small, but on a large picture it is more noticeable. If I'm not mistaken, nearly all roads were covered, except some parts of large roads and the very thin ones which requires parameter tuning of filters / algorithms used.\r\n\r\n![enter image description here][11]\r\n\r\n\r\n  [1]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\r\n  [2]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5583/kuwahara_filter.jpg\r\n  [3]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5575/connected_regions.jpg\r\n  [4]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5580/within_distance_masking.jpg\r\n  [5]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5578/skeleton_mask.jpg\r\n  [6]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5579/suliman_edge.jpg\r\n  [7]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5581/gradient_skeleton.jpg\r\n  [8]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5576/gradient_direction.jpg\r\n  [9]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5577/final.jpg\r\n  [10]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\r\n  [11]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5585/stack.gif",
      "votes": null
    },
    {
      "id": "150418",
      "postDate": "12/15/2016 03:30:40",
      "content": "<p>The rule of thumb is: if you receive some new (unseen) data tomorrow, and the steps you apply can be automated without dependency on human labeling or human judgement, it's allowed. </p>\n\n<p>In other words, doesn't have to use machine learning algorithms. </p>",
      "rawMarkdown": "The rule of thumb is: if you receive some new (unseen) data tomorrow, and the steps you apply can be automated without dependency on human labeling or human judgement, it's allowed. \r\n\r\nIn other words, doesn't have to use machine learning algorithms.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 150418,
      "author_name": "wendykan",
      "author_url": "",
      "post_date": "12/15/2016 03:30:40",
      "content": "<p>The rule of thumb is: if you receive some new (unseen) data tomorrow, and the steps you apply can be automated without dependency on human labeling or human judgement, it's allowed. </p>\n\n<p>In other words, doesn't have to use machine learning algorithms. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "150403": "I'm just wondering if manuals step are allowed? (they could be programmed, but it takes a while to code them).\r\n\r\nIt doesn't seem to violate the rules:\r\n\r\n> Submissions may not use or incorporate information from hand labeling\r\n> or human prediction of the validation dataset or test data records.\r\n\r\nBut is not receivable in case of a winning solution, right?:\r\n\r\n> The delivered software code must be capable of generating the winning\r\n> Submission and contain a description of resources required to build\r\n> and/or run the executable code successfully.\r\n\r\nFor instance taking the base picture, I want the roads:\r\n\r\n![enter image description here][1]\r\n\r\nI can apply a Kuwahara filter, which makes the roads much easier to discern.\r\n\r\n![enter image description here][2]\r\n\r\nThen I transform to 32-bit colors, and perform unsupervised connection region computation:\r\n\r\n![enter image description here][3]\r\n\r\nPicking back the 32-bit color picture, I convert back to 8-bit, and perform masking by nearby points algorithms:\r\n\r\n![enter image description here][4]\r\n\r\nTurned to a skeleton:\r\n\r\n![enter image description here][5]\r\n\r\nI can also use Suliman's fuzzy edge detection filter:\r\n\r\n![enter image description here][6]\r\n\r\nBringing back visual using gradient direction:\r\n\r\n![enter image description here][7]\r\n\r\nRoads are much easier to notice here with a skeleton:\r\n\r\n![enter image description here][8]\r\n\r\nJust a synthesis picture after inverting the skeleton, masking the kuwahara+connected regions, and stacking these 2 layers with the original pic:\r\n\r\n![enter image description here][9]![enter image description here][10]\r\n\r\nThe level of details is low here because the picture is small, but on a large picture it is more noticeable. If I'm not mistaken, nearly all roads were covered, except some parts of large roads and the very thin ones which requires parameter tuning of filters / algorithms used.\r\n\r\n![enter image description here][11]\r\n\r\n\r\n  [1]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\r\n  [2]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5583/kuwahara_filter.jpg\r\n  [3]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5575/connected_regions.jpg\r\n  [4]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5580/within_distance_masking.jpg\r\n  [5]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5578/skeleton_mask.jpg\r\n  [6]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5579/suliman_edge.jpg\r\n  [7]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5581/gradient_skeleton.jpg\r\n  [8]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5576/gradient_direction.jpg\r\n  [9]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5577/final.jpg\r\n  [10]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5582/6100-Frag_3.png\r\n  [11]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5585/stack.gif",
    "150418": "The rule of thumb is: if you receive some new (unseen) data tomorrow, and the steps you apply can be automated without dependency on human labeling or human judgement, it's allowed. \r\n\r\nIn other words, doesn't have to use machine learning algorithms."
  },
  "source": "meta"
}