{
  "id": 69941,
  "title": "Are scSE blocks are of any good in this problem? ",
  "url": "/competitions/airbus-ship-detection/discussion/69941",
  "author_name": "",
  "post_date": "2018-10-29T05:28:44.360097500Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Are scSE blocks helping anyone's score? I used them with the decoder blocks but seem no significant improvement. Can someone share their experience with it? It would be helpful. </p>",
  "messages": [
    {
      "id": "411844",
      "postDate": "10/29/2018 05:28:44",
      "content": "<p>Are scSE blocks helping anyone's score? I used them with the decoder blocks but seem no significant improvement. Can someone share their experience with it? It would be helpful. </p>",
      "rawMarkdown": "Are scSE blocks helping anyone's score? I used them with the decoder blocks but seem no significant improvement. Can someone share their experience with it? It would be helpful.",
      "votes": null
    },
    {
      "id": "412065",
      "postDate": "10/29/2018 14:19:38",
      "content": "<p>According to my experience from TGS, scSE works well if:</p>\n\n<ol>\n<li><p>the segmented object is large enough</p></li>\n<li><p>(purely empirical) used with hypercolumn</p></li>\n<li><p>Architecture is good enough + reasonable loss</p></li>\n</ol>",
      "rawMarkdown": "According to my experience from TGS, scSE works well if:\n\n1.  the segmented object is large enough\n\n2. (purely empirical) used with hypercolumn\n\n3. Architecture is good enough + reasonable loss",
      "votes": null
    },
    {
      "id": "412087",
      "postDate": "10/29/2018 14:53:46",
      "content": "<p>So since the objects in the images are small they don't seem to give an appreciable results, that's how I can see it. In this case I think it'll work well with smaller resolution images  since the ships will be bigger compared to larger resolution.\nSo when using with Unet is batch norm really required cause in the paper they've mentioned that they didn't used any batch norm, don't know why maybe because of experimental results.\nLovasz loss seems to be doing good when using scSE as many have mentioned.\nWould appreciate if some elaboration could be given about its usage.</p>",
      "rawMarkdown": "So since the objects in the images are small they don't seem to give an appreciable results, that's how I can see it. In this case I think it'll work well with smaller resolution images  since the ships will be bigger compared to larger resolution.\nSo when using with Unet is batch norm really required cause in the paper they've mentioned that they didn't used any batch norm, don't know why maybe because of experimental results.\nLovasz loss seems to be doing good when using scSE as many have mentioned.\nWould appreciate if some elaboration could be given about its usage.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 412065,
      "author_name": "alexanderliao",
      "author_url": "",
      "post_date": "10/29/2018 14:19:38",
      "content": "<p>According to my experience from TGS, scSE works well if:</p>\n\n<ol>\n<li><p>the segmented object is large enough</p></li>\n<li><p>(purely empirical) used with hypercolumn</p></li>\n<li><p>Architecture is good enough + reasonable loss</p></li>\n</ol>",
      "votes": null,
      "replies": [
        {
          "id": 412087,
          "author_name": "manu845",
          "author_url": "",
          "post_date": "10/29/2018 14:53:46",
          "content": "<p>So since the objects in the images are small they don't seem to give an appreciable results, that's how I can see it. In this case I think it'll work well with smaller resolution images  since the ships will be bigger compared to larger resolution.\nSo when using with Unet is batch norm really required cause in the paper they've mentioned that they didn't used any batch norm, don't know why maybe because of experimental results.\nLovasz loss seems to be doing good when using scSE as many have mentioned.\nWould appreciate if some elaboration could be given about its usage.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "411844": "Are scSE blocks helping anyone's score? I used them with the decoder blocks but seem no significant improvement. Can someone share their experience with it? It would be helpful.",
    "412065": "According to my experience from TGS, scSE works well if:\n\n1.  the segmented object is large enough\n\n2. (purely empirical) used with hypercolumn\n\n3. Architecture is good enough + reasonable loss",
    "412087": "So since the objects in the images are small they don't seem to give an appreciable results, that's how I can see it. In this case I think it'll work well with smaller resolution images  since the ships will be bigger compared to larger resolution.\nSo when using with Unet is batch norm really required cause in the paper they've mentioned that they didn't used any batch norm, don't know why maybe because of experimental results.\nLovasz loss seems to be doing good when using scSE as many have mentioned.\nWould appreciate if some elaboration could be given about its usage."
  },
  "source": "meta"
}