{
  "id": 27336,
  "title": "Only 1 image with waterways in the training set!!!",
  "url": "/competitions/dstl-satellite-imagery-feature-detection/discussion/27336",
  "author_name": "",
  "post_date": "2017-01-05T10:46:10.470Z",
  "votes": 2,
  "comment_count": 4,
  "views": 471,
  "content": "<p>There is only one image with waterways in the training data, can more images with waterways be added? This is a bit extreme, why  are there so few training samples? I recall in an earlier post there were no images with waterways at the start of the competition and 1 image was added to correct this but I think more should be added. What is point of limiting the train data to such an extent?</p>",
  "messages": [
    {
      "id": "154234",
      "postDate": "01/05/2017 10:46:10",
      "content": "<p>There is only one image with waterways in the training data, can more images with waterways be added? This is a bit extreme, why  are there so few training samples? I recall in an earlier post there were no images with waterways at the start of the competition and 1 image was added to correct this but I think more should be added. What is point of limiting the train data to such an extent?</p>",
      "rawMarkdown": "There is only one image with waterways in the training data, can more images with waterways be added? This is a bit extreme, why  are there so few training samples? I recall in an earlier post there were no images with waterways at the start of the competition and 1 image was added to correct this but I think more should be added. What is point of limiting the train data to such an extent?",
      "votes": null
    },
    {
      "id": "154261",
      "postDate": "01/05/2017 13:23:50",
      "content": "<p>Well, you can always crop it :D I am not there yet, so joking ;).</p>\n\n<p>However, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands.  I didn't try it myself but it is fesable. And whatever ML/DL/AI you do, you can't beat the physical phenomena.</p>",
      "rawMarkdown": "Well, you can always crop it :D I am not there yet, so joking ;).\r\n\r\nHowever, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands.  I didn't try it myself but it is fesable. And whatever ML/DL/AI you do, you can't beat the physical phenomena.",
      "votes": null
    },
    {
      "id": "154263",
      "postDate": "01/05/2017 13:29:38",
      "content": "<p>There are 2 of them:</p>\n\n<p>Class 7 Exists: ['6070_2_3', '6100_2_2']</p>\n\n<p>Ensure that you use <strong>train_wkt_v4.csv</strong></p>",
      "rawMarkdown": "There are 2 of them:\r\n\r\nClass 7 Exists: ['6070_2_3', '6100_2_2']\r\n\r\nEnsure that you use **train_wkt_v4.csv**",
      "votes": null
    },
    {
      "id": "154376",
      "postDate": "01/05/2017 23:51:09",
      "content": "<p>[quote=visoft;154261]</p>\n\n<p>However, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands. <br>\n[/quote]</p>\n\n<p>Can anyone point me toward literature that will show example of how it may be done?</p>",
      "rawMarkdown": "[quote=visoft;154261]\r\n\r\nHowever, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands.  \r\n[/quote]\r\n\r\nCan anyone point me toward literature that will show example of how it may be done?",
      "votes": null
    },
    {
      "id": "154377",
      "postDate": "01/05/2017 23:56:49",
      "content": "<p>Yes, I wouldn't say trivial but it was the easiest class for me so far. NDVI for water and some additional logic to detect foam/currents (the white parts inside waterways). I'll try to post something after I solve my issue with invalid submission.</p>",
      "rawMarkdown": "Yes, I wouldn't say trivial but it was the easiest class for me so far. NDVI for water and some additional logic to detect foam/currents (the white parts inside waterways). I'll try to post something after I solve my issue with invalid submission.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 154261,
      "author_name": "visoft",
      "author_url": "",
      "post_date": "01/05/2017 13:23:50",
      "content": "<p>Well, you can always crop it :D I am not there yet, so joking ;).</p>\n\n<p>However, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands.  I didn't try it myself but it is fesable. And whatever ML/DL/AI you do, you can't beat the physical phenomena.</p>",
      "votes": null,
      "replies": [
        {
          "id": 154376,
          "author_name": "iglovikov",
          "author_url": "",
          "post_date": "01/05/2017 23:51:09",
          "content": "<p>[quote=visoft;154261]</p>\n\n<p>However, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands. <br>\n[/quote]</p>\n\n<p>Can anyone point me toward literature that will show example of how it may be done?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 154263,
      "author_name": "zfturbo",
      "author_url": "",
      "post_date": "01/05/2017 13:29:38",
      "content": "<p>There are 2 of them:</p>\n\n<p>Class 7 Exists: ['6070_2_3', '6100_2_2']</p>\n\n<p>Ensure that you use <strong>train_wkt_v4.csv</strong></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 154377,
      "author_name": "aamaia",
      "author_url": "",
      "post_date": "01/05/2017 23:56:49",
      "content": "<p>Yes, I wouldn't say trivial but it was the easiest class for me so far. NDVI for water and some additional logic to detect foam/currents (the white parts inside waterways). I'll try to post something after I solve my issue with invalid submission.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "154234": "There is only one image with waterways in the training data, can more images with waterways be added? This is a bit extreme, why  are there so few training samples? I recall in an earlier post there were no images with waterways at the start of the competition and 1 image was added to correct this but I think more should be added. What is point of limiting the train data to such an extent?",
    "154261": "Well, you can always crop it :D I am not there yet, so joking ;).\r\n\r\nHowever, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands.  I didn't try it myself but it is fesable. And whatever ML/DL/AI you do, you can't beat the physical phenomena.",
    "154263": "There are 2 of them:\r\n\r\nClass 7 Exists: ['6070_2_3', '6100_2_2']\r\n\r\nEnsure that you use **train_wkt_v4.csv**",
    "154376": "[quote=visoft;154261]\r\n\r\nHowever, in some threads (amaia I think) told that detecting trees, and water is trivial because of the infrared bands.  \r\n[/quote]\r\n\r\nCan anyone point me toward literature that will show example of how it may be done?",
    "154377": "Yes, I wouldn't say trivial but it was the easiest class for me so far. NDVI for water and some additional logic to detect foam/currents (the white parts inside waterways). I'll try to post something after I solve my issue with invalid submission."
  },
  "source": "meta"
}