{
  "id": 291532,
  "title": "Simple network for semantic segmentation",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/291532",
  "author_name": "Gere",
  "post_date": "2021-11-29T21:25:18.719000",
  "votes": 8,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I'm trying to get a feel for segmentation networks and so I wrote a test with a very simple toy network: <a href=\"https://www.kaggle.com/superant/simple-nn-segmentation\" target=\"_blank\">https://www.kaggle.com/superant/simple-nn-segmentation</a><br>\nHowever, the current network is just a random guess. For simplicity it only does semantic segmentation (no instance splitting).</p>\n<p>I'd appreciate advice or guesses on the following questions:</p>\n<ol>\n<li>Do you have an idea which simple architecture could work better? (which direction to try for channel numbers, kernel sizes, other layer types - it shouldn't get much more complex though)</li>\n<li>Is it a good idea to keep the image size (W/H) constant across the whole network?</li>\n<li>Any other simple improvement that may improve score without increasing computational complexity too much?</li>\n<li>Could the output of such a network theoretically be used as the backbone in Detectron2?</li>\n</ol>",
  "messages": [
    {
      "id": 1599820,
      "postDate": "2021-11-29T21:25:18.720Z",
      "content": "<p>I'm trying to get a feel for segmentation networks and so I wrote a test with a very simple toy network: <a href=\"https://www.kaggle.com/superant/simple-nn-segmentation\" target=\"_blank\">https://www.kaggle.com/superant/simple-nn-segmentation</a><br>\nHowever, the current network is just a random guess. For simplicity it only does semantic segmentation (no instance splitting).</p>\n<p>I'd appreciate advice or guesses on the following questions:</p>\n<ol>\n<li>Do you have an idea which simple architecture could work better? (which direction to try for channel numbers, kernel sizes, other layer types - it shouldn't get much more complex though)</li>\n<li>Is it a good idea to keep the image size (W/H) constant across the whole network?</li>\n<li>Any other simple improvement that may improve score without increasing computational complexity too much?</li>\n<li>Could the output of such a network theoretically be used as the backbone in Detectron2?</li>\n</ol>",
      "rawMarkdown": "I'm trying to get a feel for segmentation networks and so I wrote a test with a very simple toy network: https://www.kaggle.com/superant/simple-nn-segmentation\nHowever, the current network is just a random guess. For simplicity it only does semantic segmentation (no instance splitting).\n\nI'd appreciate advice or guesses on the following questions:\n1. Do you have an idea which simple architecture could work better? (which direction to try for channel numbers, kernel sizes, other layer types - it shouldn't get much more complex though)\n2. Is it a good idea to keep the image size (W/H) constant across the whole network?\n3. Any other simple improvement that may improve score without increasing computational complexity too much?\n4. Could the output of such a network theoretically be used as the backbone in Detectron2?",
      "votes": 8
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1599820": "I'm trying to get a feel for segmentation networks and so I wrote a test with a very simple toy network: https://www.kaggle.com/superant/simple-nn-segmentation\nHowever, the current network is just a random guess. For simplicity it only does semantic segmentation (no instance splitting).\n\nI'd appreciate advice or guesses on the following questions:\n1. Do you have an idea which simple architecture could work better? (which direction to try for channel numbers, kernel sizes, other layer types - it shouldn't get much more complex though)\n2. Is it a good idea to keep the image size (W/H) constant across the whole network?\n3. Any other simple improvement that may improve score without increasing computational complexity too much?\n4. Could the output of such a network theoretically be used as the backbone in Detectron2?"
  }
}