{
  "id": 34671,
  "title": "Binary tile classification",
  "url": "/competitions/noaa-fisheries-steller-sea-lion-population-count/discussion/34671",
  "author_name": "",
  "post_date": "2017-06-13T22:55:32.879229300Z",
  "votes": 4,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Anyone performing binary tile classification as sort of preprocessing step?</p>\n\n<p>Results are not perfect and code is pretty dirty now, but I will clean it and post in few days on github.</p>\n\n<p>Here are some results:</p>\n\n<pre><code>Train\n80 images\nNon zero samples 5452 / 24394\nVal\n9 images\nNon zero samples 216 / 2316\nTP: 169\nFP: 19\nTN: 2081\nFN: 47\nPrecision: 0.898936170213\nRecall: 0.782407407407\nAccuracy: 0.971502590674\n</code></pre>\n\n<p>My val images were: <code>image_ids_val= [1,10,100,300,400,500,600,700,900]</code></p>",
  "messages": [
    {
      "id": "192532",
      "postDate": "06/13/2017 22:55:32",
      "content": "<p>Anyone performing binary tile classification as sort of preprocessing step?</p>\n\n<p>Results are not perfect and code is pretty dirty now, but I will clean it and post in few days on github.</p>\n\n<p>Here are some results:</p>\n\n<pre><code>Train\n80 images\nNon zero samples 5452 / 24394\nVal\n9 images\nNon zero samples 216 / 2316\nTP: 169\nFP: 19\nTN: 2081\nFN: 47\nPrecision: 0.898936170213\nRecall: 0.782407407407\nAccuracy: 0.971502590674\n</code></pre>\n\n<p>My val images were: <code>image_ids_val= [1,10,100,300,400,500,600,700,900]</code></p>",
      "rawMarkdown": "Anyone performing binary tile classification as sort of preprocessing step?\n\nResults are not perfect and code is pretty dirty now, but I will clean it and post in few days on github.\n\nHere are some results:\n\n    Train\n    80 images\n    Non zero samples 5452 / 24394\n    Val\n    9 images\n    Non zero samples 216 / 2316\n    TP: 169\n    FP: 19\n    TN: 2081\n    FN: 47\n    Precision: 0.898936170213\n    Recall: 0.782407407407\n    Accuracy: 0.971502590674\n\nMy val images were: `image_ids_val= [1,10,100,300,400,500,600,700,900]`",
      "votes": null
    },
    {
      "id": "192978",
      "postDate": "06/15/2017 09:07:46",
      "content": "<p>Hi. Nice work!\nI'm trying something similar except it's rather difficult, using this approach, to handle multiple sealions hanging out in a single tile. How do you handle this?</p>",
      "rawMarkdown": "Hi. Nice work!\nI'm trying something similar except it's rather difficult, using this approach, to handle multiple sealions hanging out in a single tile. How do you handle this?",
      "votes": null
    },
    {
      "id": "193188",
      "postDate": "06/15/2017 19:32:19",
      "content": "<p>It's just a binary classifier and you need another one to predict count of sea lions on tile.</p>",
      "rawMarkdown": "It's just a binary classifier and you need another one to predict count of sea lions on tile.",
      "votes": null
    },
    {
      "id": "193633",
      "postDate": "06/17/2017 08:48:58",
      "content": "<p>Nice idea. You may use feature extrator and aggregate them to possibly obtain a count predictor.</p>",
      "rawMarkdown": "Nice idea. You may use feature extrator and aggregate them to possibly obtain a count predictor.",
      "votes": null
    },
    {
      "id": "193689",
      "postDate": "06/17/2017 15:09:32",
      "content": "<p>I think this approach may help us to reduce calculation cost.<br>\nIf we remove area where there is no sea-lions obviously, we can focus on important part.</p>",
      "rawMarkdown": "I think this approach may help us to reduce calculation cost.<br>\nIf we remove area where there is no sea-lions obviously, we can focus on important part.",
      "votes": null
    },
    {
      "id": "194449",
      "postDate": "06/20/2017 16:42:05",
      "content": "<p>Visually inspecting results of binary classification on test set I found some typical problems:</p>\n\n<ol>\n<li>700_debug.jpg Not sure if they are sea lion at all.</li>\n<li>2000_debug.jpg Scale problem.</li>\n<li>5100_debug.jpg Scale problem.</li>\n<li>10200_debug.jpg Not sure if they are sea lions and they look like a rocks.</li>\n</ol>",
      "rawMarkdown": "Visually inspecting results of binary classification on test set I found some typical problems:\n\n1. 700_debug.jpg Not sure if they are sea lion at all.\n2. 2000_debug.jpg Scale problem.\n3. 5100_debug.jpg Scale problem.\n4. 10200_debug.jpg Not sure if they are sea lions and they look like a rocks.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 192978,
      "author_name": "sanderdalm",
      "author_url": "",
      "post_date": "06/15/2017 09:07:46",
      "content": "<p>Hi. Nice work!\nI'm trying something similar except it's rather difficult, using this approach, to handle multiple sealions hanging out in a single tile. How do you handle this?</p>",
      "votes": null,
      "replies": [
        {
          "id": 193188,
          "author_name": "mrgloom",
          "author_url": "",
          "post_date": "06/15/2017 19:32:19",
          "content": "<p>It's just a binary classifier and you need another one to predict count of sea lions on tile.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 193633,
      "author_name": "captaindadja",
      "author_url": "",
      "post_date": "06/17/2017 08:48:58",
      "content": "<p>Nice idea. You may use feature extrator and aggregate them to possibly obtain a count predictor.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 193689,
      "author_name": "toshik",
      "author_url": "",
      "post_date": "06/17/2017 15:09:32",
      "content": "<p>I think this approach may help us to reduce calculation cost.<br>\nIf we remove area where there is no sea-lions obviously, we can focus on important part.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 194449,
      "author_name": "mrgloom",
      "author_url": "",
      "post_date": "06/20/2017 16:42:05",
      "content": "<p>Visually inspecting results of binary classification on test set I found some typical problems:</p>\n\n<ol>\n<li>700_debug.jpg Not sure if they are sea lion at all.</li>\n<li>2000_debug.jpg Scale problem.</li>\n<li>5100_debug.jpg Scale problem.</li>\n<li>10200_debug.jpg Not sure if they are sea lions and they look like a rocks.</li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "192532": "Anyone performing binary tile classification as sort of preprocessing step?\n\nResults are not perfect and code is pretty dirty now, but I will clean it and post in few days on github.\n\nHere are some results:\n\n    Train\n    80 images\n    Non zero samples 5452 / 24394\n    Val\n    9 images\n    Non zero samples 216 / 2316\n    TP: 169\n    FP: 19\n    TN: 2081\n    FN: 47\n    Precision: 0.898936170213\n    Recall: 0.782407407407\n    Accuracy: 0.971502590674\n\nMy val images were: `image_ids_val= [1,10,100,300,400,500,600,700,900]`",
    "192978": "Hi. Nice work!\nI'm trying something similar except it's rather difficult, using this approach, to handle multiple sealions hanging out in a single tile. How do you handle this?",
    "193188": "It's just a binary classifier and you need another one to predict count of sea lions on tile.",
    "193633": "Nice idea. You may use feature extrator and aggregate them to possibly obtain a count predictor.",
    "193689": "I think this approach may help us to reduce calculation cost.<br>\nIf we remove area where there is no sea-lions obviously, we can focus on important part.",
    "194449": "Visually inspecting results of binary classification on test set I found some typical problems:\n\n1. 700_debug.jpg Not sure if they are sea lion at all.\n2. 2000_debug.jpg Scale problem.\n3. 5100_debug.jpg Scale problem.\n4. 10200_debug.jpg Not sure if they are sea lions and they look like a rocks."
  },
  "source": "meta"
}