{
  "id": 69841,
  "title": "Some ideas on separating ships that are close together",
  "url": "/competitions/airbus-ship-detection/discussion/69841",
  "author_name": "",
  "post_date": "2018-10-27T22:33:22.033509Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<ol>\n<li><p>Watershed</p></li>\n<li><p><a href=\"https://arxiv.org/pdf/1801.01315.pdf\">PixelLink</a></p></li>\n<li><p><a href=\"https://arxiv.org/pdf/1704.04503.pdf\">Soft-NMS</a></p></li>\n</ol>\n\n<p>....</p>",
  "messages": [
    {
      "id": "411346",
      "postDate": "10/27/2018 22:33:22",
      "content": "<ol>\n<li><p>Watershed</p></li>\n<li><p><a href=\"https://arxiv.org/pdf/1801.01315.pdf\">PixelLink</a></p></li>\n<li><p><a href=\"https://arxiv.org/pdf/1704.04503.pdf\">Soft-NMS</a></p></li>\n</ol>\n\n<p>....</p>",
      "rawMarkdown": "1. Watershed\n\n2. [PixelLink][1]\n\n3. [Soft-NMS][2]\n\n....\n\n\n  [1]: https://arxiv.org/pdf/1801.01315.pdf\n  [2]: https://arxiv.org/pdf/1704.04503.pdf",
      "votes": null
    },
    {
      "id": "411923",
      "postDate": "10/29/2018 07:59:08",
      "content": "<p>Hi there @Alexander Liao.</p>\n\n<p>Looking at the results of our latest model (where we didn't do anything to fix the separation of ships)</p>\n\n<pre><code>Empty | f2: 0.994 | gain: 0.003\nNon Empty f2: 0.422 | gain: 0.277\n1 ship f2: 0.445 | gain: 0.178\n2-5 ships f2: 0.393 | gain: 0.083\n5-10 ships f2: 0.308 | gain: 0.012\n10+ ships f2: 0.205 | gain: 0.005\n</code></pre>\n\n<p>one can see that the separation is way less important than not missing single ships.\nWe are thinking about how to tweak the loss and target to do just that.</p>\n\n<p>If you want to go down the road of separating ships, what worked well for us in the past was employing the original UNet loss. Full write-up and code (in pytorch) is available <a href=\"https://github.com/neptune-ml/open-solution-mapping-challenge\">here</a>.</p>\n\n<p>Great job on the Salt btw!</p>",
      "rawMarkdown": "Hi there @Alexander Liao.\n\nLooking at the results of our latest model (where we didn't do anything to fix the separation of ships)\n\n    Empty | f2: 0.994 | gain: 0.003\n    Non Empty f2: 0.422 | gain: 0.277\n    1 ship f2: 0.445 | gain: 0.178\n    2-5 ships f2: 0.393 | gain: 0.083\n    5-10 ships f2: 0.308 | gain: 0.012\n    10+ ships f2: 0.205 | gain: 0.005\n\none can see that the separation is way less important than not missing single ships.\nWe are thinking about how to tweak the loss and target to do just that.\n\nIf you want to go down the road of separating ships, what worked well for us in the past was employing the original UNet loss. Full write-up and code (in pytorch) is available [here](https://github.com/neptune-ml/open-solution-mapping-challenge).\n\nGreat job on the Salt btw!",
      "votes": null
    },
    {
      "id": "413897",
      "postDate": "11/01/2018 18:25:59",
      "content": "<p>Thanks!</p>",
      "rawMarkdown": "Thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 411923,
      "author_name": "jakubczakon",
      "author_url": "",
      "post_date": "10/29/2018 07:59:08",
      "content": "<p>Hi there @Alexander Liao.</p>\n\n<p>Looking at the results of our latest model (where we didn't do anything to fix the separation of ships)</p>\n\n<pre><code>Empty | f2: 0.994 | gain: 0.003\nNon Empty f2: 0.422 | gain: 0.277\n1 ship f2: 0.445 | gain: 0.178\n2-5 ships f2: 0.393 | gain: 0.083\n5-10 ships f2: 0.308 | gain: 0.012\n10+ ships f2: 0.205 | gain: 0.005\n</code></pre>\n\n<p>one can see that the separation is way less important than not missing single ships.\nWe are thinking about how to tweak the loss and target to do just that.</p>\n\n<p>If you want to go down the road of separating ships, what worked well for us in the past was employing the original UNet loss. Full write-up and code (in pytorch) is available <a href=\"https://github.com/neptune-ml/open-solution-mapping-challenge\">here</a>.</p>\n\n<p>Great job on the Salt btw!</p>",
      "votes": null,
      "replies": [
        {
          "id": 413897,
          "author_name": "alexanderliao",
          "author_url": "",
          "post_date": "11/01/2018 18:25:59",
          "content": "<p>Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "411346": "1. Watershed\n\n2. [PixelLink][1]\n\n3. [Soft-NMS][2]\n\n....\n\n\n  [1]: https://arxiv.org/pdf/1801.01315.pdf\n  [2]: https://arxiv.org/pdf/1704.04503.pdf",
    "411923": "Hi there @Alexander Liao.\n\nLooking at the results of our latest model (where we didn't do anything to fix the separation of ships)\n\n    Empty | f2: 0.994 | gain: 0.003\n    Non Empty f2: 0.422 | gain: 0.277\n    1 ship f2: 0.445 | gain: 0.178\n    2-5 ships f2: 0.393 | gain: 0.083\n    5-10 ships f2: 0.308 | gain: 0.012\n    10+ ships f2: 0.205 | gain: 0.005\n\none can see that the separation is way less important than not missing single ships.\nWe are thinking about how to tweak the loss and target to do just that.\n\nIf you want to go down the road of separating ships, what worked well for us in the past was employing the original UNet loss. Full write-up and code (in pytorch) is available [here](https://github.com/neptune-ml/open-solution-mapping-challenge).\n\nGreat job on the Salt btw!",
    "413897": "Thanks!"
  },
  "source": "meta"
}