{
  "id": 35397,
  "title": "Window level regression",
  "url": "/competitions/noaa-fisheries-steller-sea-lion-population-count/discussion/35397",
  "author_name": "",
  "post_date": "2017-06-27T20:18:02.890601100Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi all,</p>\n\n<p>For anyone interested here's an approach that extends mrgloom's image-level regressions script to work at the window level.</p>\n\n<p>It extracts 512x512 windows from the images, counts the lions using the dots and blob detection, and compiles a significantly enlarged training set. </p>\n\n<p>Prediction is done using a sliding window. Any incomplete edges are dropped, which isn't ideal</p>\n\n<p>It's not totally finished, never really scored that well, and won't be as interesting as I'm sure the winning solutions will be, but I think it's different to the other approaches posted so far (?). </p>\n\n<p>Enjoy!</p>\n\n<p><a href=\"https://github.com/garethjns/Kaggle-Sea-Lions-Solution/tree/master\">https://github.com/garethjns/Kaggle-Sea-Lions-Solution/tree/master</a></p>",
  "messages": [
    {
      "id": "196623",
      "postDate": "06/27/2017 20:18:02",
      "content": "<p>Hi all,</p>\n\n<p>For anyone interested here's an approach that extends mrgloom's image-level regressions script to work at the window level.</p>\n\n<p>It extracts 512x512 windows from the images, counts the lions using the dots and blob detection, and compiles a significantly enlarged training set. </p>\n\n<p>Prediction is done using a sliding window. Any incomplete edges are dropped, which isn't ideal</p>\n\n<p>It's not totally finished, never really scored that well, and won't be as interesting as I'm sure the winning solutions will be, but I think it's different to the other approaches posted so far (?). </p>\n\n<p>Enjoy!</p>\n\n<p><a href=\"https://github.com/garethjns/Kaggle-Sea-Lions-Solution/tree/master\">https://github.com/garethjns/Kaggle-Sea-Lions-Solution/tree/master</a></p>",
      "rawMarkdown": "Hi all,\n\nFor anyone interested here's an approach that extends mrgloom's image-level regressions script to work at the window level.\n\nIt extracts 512x512 windows from the images, counts the lions using the dots and blob detection, and compiles a significantly enlarged training set. \n\nPrediction is done using a sliding window. Any incomplete edges are dropped, which isn't ideal\n\nIt's not totally finished, never really scored that well, and won't be as interesting as I'm sure the winning solutions will be, but I think it's different to the other approaches posted so far (?). \n\nEnjoy!\n\nhttps://github.com/garethjns/Kaggle-Sea-Lions-Solution/tree/master",
      "votes": null
    },
    {
      "id": "196679",
      "postDate": "06/27/2017 22:44:48",
      "content": "<p>You should have taken a deeper network!</p>",
      "rawMarkdown": "You should have taken a deeper network!",
      "votes": null
    },
    {
      "id": "196709",
      "postDate": "06/28/2017 01:07:17",
      "content": "<p>Looking forward to your solution. :P</p>",
      "rawMarkdown": "Looking forward to your solution. :P",
      "votes": null
    },
    {
      "id": "196794",
      "postDate": "06/28/2017 05:47:58",
      "content": "<p>Yeah, I did want to try VGG16, but wouldn't have had time to run it on my Geforce 960! Congratulations on your top ten finish.</p>",
      "rawMarkdown": "Yeah, I did want to try VGG16, but wouldn't have had time to run it on my Geforce 960! Congratulations on your top ten finish.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 196679,
      "author_name": "asanakoev",
      "author_url": "",
      "post_date": "06/27/2017 22:44:48",
      "content": "<p>You should have taken a deeper network!</p>",
      "votes": null,
      "replies": [
        {
          "id": 196709,
          "author_name": "badoun",
          "author_url": "",
          "post_date": "06/28/2017 01:07:17",
          "content": "<p>Looking forward to your solution. :P</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 196794,
          "author_name": "garethjns",
          "author_url": "",
          "post_date": "06/28/2017 05:47:58",
          "content": "<p>Yeah, I did want to try VGG16, but wouldn't have had time to run it on my Geforce 960! Congratulations on your top ten finish.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "196623": "Hi all,\n\nFor anyone interested here's an approach that extends mrgloom's image-level regressions script to work at the window level.\n\nIt extracts 512x512 windows from the images, counts the lions using the dots and blob detection, and compiles a significantly enlarged training set. \n\nPrediction is done using a sliding window. Any incomplete edges are dropped, which isn't ideal\n\nIt's not totally finished, never really scored that well, and won't be as interesting as I'm sure the winning solutions will be, but I think it's different to the other approaches posted so far (?). \n\nEnjoy!\n\nhttps://github.com/garethjns/Kaggle-Sea-Lions-Solution/tree/master",
    "196679": "You should have taken a deeper network!",
    "196709": "Looking forward to your solution. :P",
    "196794": "Yeah, I did want to try VGG16, but wouldn't have had time to run it on my Geforce 960! Congratulations on your top ten finish."
  },
  "source": "meta"
}