{
  "id": 30123,
  "title": "9th place solution",
  "url": "/competitions/dstl-satellite-imagery-feature-detection/writeups/toshi-k-9th-place-solution",
  "author_name": "",
  "post_date": "2017-03-16T03:01:49.997Z",
  "votes": 24,
  "comment_count": 8,
  "views": 0,
  "content": "<p>Hi Everyone,</p>\n\n<p>Ending final validation, I share my solution.<br>\nI use 3 band images and SWIR images.<br>\nMain approach is Deep Learning based on SegNet (for class 1~6, 8~10).<br>\nI also use V.Osin's kaggle script (for class 7).</p>\n\n<p>Code: <a href=\"https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection\">https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection</a></p>\n\n<p><img src=\"https://raw.githubusercontent.com/toshi-k/kaggle-satellite-imagery-feature-detection/master/img/solution.png\" alt=\"conceptual diagram\" title=\"\"></p>",
  "messages": [
    {
      "id": "167845",
      "postDate": "03/15/2017 16:21:42",
      "content": "<p>Hi Everyone,</p>\n\n<p>Ending final validation, I share my solution.<br>\nI use 3 band images and SWIR images.<br>\nMain approach is Deep Learning based on SegNet (for class 1~6, 8~10).<br>\nI also use V.Osin's kaggle script (for class 7).</p>\n\n<p>Code: <a href=\"https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection\">https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection</a></p>\n\n<p><img src=\"https://raw.githubusercontent.com/toshi-k/kaggle-satellite-imagery-feature-detection/master/img/solution.png\" alt=\"conceptual diagram\" title=\"\"></p>",
      "rawMarkdown": "Hi Everyone,\n\nEnding final validation, I share my solution.<br>\nI use 3 band images and SWIR images.<br>\nMain approach is Deep Learning based on SegNet (for class 1~6, 8~10).<br>\nI also use V.Osin's kaggle script (for class 7).\n\nCode: https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection\n\n![conceptual diagram][1]\n\n\n  [1]: https://raw.githubusercontent.com/toshi-k/kaggle-satellite-imagery-feature-detection/master/img/solution.png",
      "votes": null
    },
    {
      "id": "167848",
      "postDate": "03/15/2017 16:29:05",
      "content": "<p>Thank you for sharing and congrats for your possition.</p>",
      "rawMarkdown": "Thank you for sharing and congrats for your possition.",
      "votes": null
    },
    {
      "id": "167853",
      "postDate": "03/15/2017 16:54:29",
      "content": "<p>Thanks Toshi_K for sharing. Congratulations!</p>",
      "rawMarkdown": "Thanks Toshi_K for sharing. Congratulations!",
      "votes": null
    },
    {
      "id": "167860",
      "postDate": "03/15/2017 17:19:09",
      "content": "<p>How did you come with decision to use only 3-band and A-band, without M? </p>",
      "rawMarkdown": "How did you come with decision to use only 3-band and A-band, without M?",
      "votes": null
    },
    {
      "id": "167861",
      "postDate": "03/15/2017 17:24:21",
      "content": "<p>I think 3-band images are made from M-band. It seems to be redundant in some way.<br>\nI hear that ZFTurbo use all of them, which may improve score.<br>\n<a href=\"https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/discussion/29747\">https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/discussion/29747</a></p>",
      "rawMarkdown": "I think 3-band images are made from M-band. It seems to be redundant in some way.<br>\nI hear that ZFTurbo use all of them, which may improve score.<br>\nhttps://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/discussion/29747",
      "votes": null
    },
    {
      "id": "167866",
      "postDate": "03/15/2017 17:36:44",
      "content": "<p>So, you just haven't tried to include M-band? For sure, 3-band made from P and M but only from small part of M (3 channels of 8). I'm asking because almost everybody seems to be using M-band to get decent results and your solution shows that you can get quite high score even without M which is interesting! </p>",
      "rawMarkdown": "So, you just haven't tried to include M-band? For sure, 3-band made from P and M but only from small part of M (3 channels of 8). I'm asking because almost everybody seems to be using M-band to get decent results and your solution shows that you can get quite high score even without M which is interesting!",
      "votes": null
    },
    {
      "id": "167890",
      "postDate": "03/15/2017 18:53:13",
      "content": "<p>thanks for sharing</p>\n\n<p>did you have one model predicting all classes at once, or one model per class? (based on the sketch it's one model per class)\nalso how is SegNet different to Unet? (did you try both and did Unet perform worse?)</p>",
      "rawMarkdown": "thanks for sharing\n\ndid you have one model predicting all classes at once, or one model per class? (based on the sketch it's one model per class)\nalso how is SegNet different to Unet? (did you try both and did Unet perform worse?)",
      "votes": null
    },
    {
      "id": "167896",
      "postDate": "03/15/2017 19:13:35",
      "content": "<p>Thank you for sharing code.</p>",
      "rawMarkdown": "Thank you for sharing code.",
      "votes": null
    },
    {
      "id": "168025",
      "postDate": "03/16/2017 03:13:39",
      "content": "<p>I build models for each class. This enable to train models carefully for each class.<br>\nI didn't enough time to research about differences between SegNet and Unet.<br>\nYou can find code of my models here.<br>\n<a href=\"https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection/blob/master/source/02_class_1-6_8-10/2_model.lua\">https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection/blob/master/source/02_class_1-6_8-10/2_model.lua</a></p>",
      "rawMarkdown": "I build models for each class. This enable to train models carefully for each class.<br>\nI didn't enough time to research about differences between SegNet and Unet.<br>\nYou can find code of my models here.<br>\nhttps://github.com/toshi-k/kaggle-satellite-imagery-feature-detection/blob/master/source/02_class_1-6_8-10/2_model.lua",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 167848,
      "author_name": "santiagomota",
      "author_url": "",
      "post_date": "03/15/2017 16:29:05",
      "content": "<p>Thank you for sharing and congrats for your possition.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 167853,
      "author_name": "etakla",
      "author_url": "",
      "post_date": "03/15/2017 16:54:29",
      "content": "<p>Thanks Toshi_K for sharing. Congratulations!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 167860,
      "author_name": "ceperaang",
      "author_url": "",
      "post_date": "03/15/2017 17:19:09",
      "content": "<p>How did you come with decision to use only 3-band and A-band, without M? </p>",
      "votes": null,
      "replies": [
        {
          "id": 167861,
          "author_name": "toshik",
          "author_url": "",
          "post_date": "03/15/2017 17:24:21",
          "content": "<p>I think 3-band images are made from M-band. It seems to be redundant in some way.<br>\nI hear that ZFTurbo use all of them, which may improve score.<br>\n<a href=\"https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/discussion/29747\">https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/discussion/29747</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 167866,
          "author_name": "ceperaang",
          "author_url": "",
          "post_date": "03/15/2017 17:36:44",
          "content": "<p>So, you just haven't tried to include M-band? For sure, 3-band made from P and M but only from small part of M (3 channels of 8). I'm asking because almost everybody seems to be using M-band to get decent results and your solution shows that you can get quite high score even without M which is interesting! </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 167890,
      "author_name": "steelrose",
      "author_url": "",
      "post_date": "03/15/2017 18:53:13",
      "content": "<p>thanks for sharing</p>\n\n<p>did you have one model predicting all classes at once, or one model per class? (based on the sketch it's one model per class)\nalso how is SegNet different to Unet? (did you try both and did Unet perform worse?)</p>",
      "votes": null,
      "replies": [
        {
          "id": 168025,
          "author_name": "toshik",
          "author_url": "",
          "post_date": "03/16/2017 03:13:39",
          "content": "<p>I build models for each class. This enable to train models carefully for each class.<br>\nI didn't enough time to research about differences between SegNet and Unet.<br>\nYou can find code of my models here.<br>\n<a href=\"https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection/blob/master/source/02_class_1-6_8-10/2_model.lua\">https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection/blob/master/source/02_class_1-6_8-10/2_model.lua</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 167896,
      "author_name": "samihaq",
      "author_url": "",
      "post_date": "03/15/2017 19:13:35",
      "content": "<p>Thank you for sharing code.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "167845": "Hi Everyone,\n\nEnding final validation, I share my solution.<br>\nI use 3 band images and SWIR images.<br>\nMain approach is Deep Learning based on SegNet (for class 1~6, 8~10).<br>\nI also use V.Osin's kaggle script (for class 7).\n\nCode: https://github.com/toshi-k/kaggle-satellite-imagery-feature-detection\n\n![conceptual diagram][1]\n\n\n  [1]: https://raw.githubusercontent.com/toshi-k/kaggle-satellite-imagery-feature-detection/master/img/solution.png",
    "167848": "Thank you for sharing and congrats for your possition.",
    "167853": "Thanks Toshi_K for sharing. Congratulations!",
    "167860": "How did you come with decision to use only 3-band and A-band, without M?",
    "167861": "I think 3-band images are made from M-band. It seems to be redundant in some way.<br>\nI hear that ZFTurbo use all of them, which may improve score.<br>\nhttps://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/discussion/29747",
    "167866": "So, you just haven't tried to include M-band? For sure, 3-band made from P and M but only from small part of M (3 channels of 8). I'm asking because almost everybody seems to be using M-band to get decent results and your solution shows that you can get quite high score even without M which is interesting!",
    "167890": "thanks for sharing\n\ndid you have one model predicting all classes at once, or one model per class? (based on the sketch it's one model per class)\nalso how is SegNet different to Unet? (did you try both and did Unet perform worse?)",
    "167896": "Thank you for sharing code.",
    "168025": "I build models for each class. This enable to train models carefully for each class.<br>\nI didn't enough time to research about differences between SegNet and Unet.<br>\nYou can find code of my models here.<br>\nhttps://github.com/toshi-k/kaggle-satellite-imagery-feature-detection/blob/master/source/02_class_1-6_8-10/2_model.lua"
  },
  "source": "meta"
}