{
  "id": 206286,
  "title": "Worse LB Score with Only Cortex or Stuct Preds",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/206286",
  "author_name": "",
  "post_date": "2020-12-23T21:58:39.527063800Z",
  "votes": 8,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I assume if not all most of the gromeruli to be present in the cortex area. In few cases there might be some outside masks due to ground truth annotation mistakes either in cortex polygon or gromeruli polygons. </p>\n<p>Keeping this assumption in my mind I applied a simple post processing to my initial submission, which is to remove all predictions if it is out of the cortex region. My LB score dropped by 0.007 which quite surprised me as I was expecting an increase. </p>\n<p>How is your experience with such approach?</p>\n<p>Kernel: <a href=\"https://www.kaggle.com/keremt/hubmap-fast-submission\" target=\"_blank\">https://www.kaggle.com/keremt/hubmap-fast-submission</a></p>\n<p><strong>Update:</strong> When including all anatomical structure LB drops 0.003. This means there are mask annotations outside structure polygon.</p>",
  "messages": [
    {
      "id": "1124384",
      "postDate": "12/23/2020 21:58:39",
      "content": "<p>I assume if not all most of the gromeruli to be present in the cortex area. In few cases there might be some outside masks due to ground truth annotation mistakes either in cortex polygon or gromeruli polygons. </p>\n<p>Keeping this assumption in my mind I applied a simple post processing to my initial submission, which is to remove all predictions if it is out of the cortex region. My LB score dropped by 0.007 which quite surprised me as I was expecting an increase. </p>\n<p>How is your experience with such approach?</p>\n<p>Kernel: <a href=\"https://www.kaggle.com/keremt/hubmap-fast-submission\" target=\"_blank\">https://www.kaggle.com/keremt/hubmap-fast-submission</a></p>\n<p><strong>Update:</strong> When including all anatomical structure LB drops 0.003. This means there are mask annotations outside structure polygon.</p>",
      "rawMarkdown": "I assume if not all most of the gromeruli to be present in the cortex area. In few cases there might be some outside masks due to ground truth annotation mistakes either in cortex polygon or gromeruli polygons. \n\nKeeping this assumption in my mind I applied a simple post processing to my initial submission, which is to remove all predictions if it is out of the cortex region. My LB score dropped by 0.007 which quite surprised me as I was expecting an increase. \n\nHow is your experience with such approach?\n\nKernel: https://www.kaggle.com/keremt/hubmap-fast-submission\n\n**Update:** When including all anatomical structure LB drops 0.003. This means there are mask annotations outside structure polygon.",
      "votes": null
    },
    {
      "id": "1125718",
      "postDate": "12/25/2020 02:18:52",
      "content": "<p>I think so！My result is the same！</p>",
      "rawMarkdown": "I think so！My result is the same！",
      "votes": null
    },
    {
      "id": "1126036",
      "postDate": "12/25/2020 09:29:31",
      "content": "<p>Yep, been there. Dome that. :)</p>",
      "rawMarkdown": "Yep, been there. Dome that. :)",
      "votes": null
    },
    {
      "id": "1129898",
      "postDate": "12/28/2020 16:14:07",
      "content": "<p>my understanding is that there should be no glomeruli outside of the cortex. Due to annotation errors, I'm afraid that you are penalized for missing otherwise incorrect labels. This is just a guess. A better approach may be to exclude every other anatomical area from training.</p>",
      "rawMarkdown": "my understanding is that there should be no glomeruli outside of the cortex. Due to annotation errors, I'm afraid that you are penalized for missing otherwise incorrect labels. This is just a guess. A better approach may be to exclude every other anatomical area from training.",
      "votes": null
    },
    {
      "id": "1129905",
      "postDate": "12/28/2020 16:25:47",
      "content": "<p>There is no absolute boundary between regions in the kidney. Therefore, the AS masks provided by subject matter experts serve as a rough guide. </p>",
      "rawMarkdown": "There is no absolute boundary between regions in the kidney. Therefore, the AS masks provided by subject matter experts serve as a rough guide.",
      "votes": null
    },
    {
      "id": "1130025",
      "postDate": "12/28/2020 17:52:11",
      "content": "<p>Thanks for the clarification!</p>",
      "rawMarkdown": "Thanks for the clarification!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1125718,
      "author_name": "zzemily",
      "author_url": "",
      "post_date": "12/25/2020 02:18:52",
      "content": "<p>I think so！My result is the same！</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1126036,
      "author_name": "sakvaua",
      "author_url": "",
      "post_date": "12/25/2020 09:29:31",
      "content": "<p>Yep, been there. Dome that. :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1129898,
      "author_name": "andrasferenczi",
      "author_url": "",
      "post_date": "12/28/2020 16:14:07",
      "content": "<p>my understanding is that there should be no glomeruli outside of the cortex. Due to annotation errors, I'm afraid that you are penalized for missing otherwise incorrect labels. This is just a guess. A better approach may be to exclude every other anatomical area from training.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1129905,
      "author_name": "leahscherschel",
      "author_url": "",
      "post_date": "12/28/2020 16:25:47",
      "content": "<p>There is no absolute boundary between regions in the kidney. Therefore, the AS masks provided by subject matter experts serve as a rough guide. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1130025,
          "author_name": "keremt",
          "author_url": "",
          "post_date": "12/28/2020 17:52:11",
          "content": "<p>Thanks for the clarification!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1124384": "I assume if not all most of the gromeruli to be present in the cortex area. In few cases there might be some outside masks due to ground truth annotation mistakes either in cortex polygon or gromeruli polygons. \n\nKeeping this assumption in my mind I applied a simple post processing to my initial submission, which is to remove all predictions if it is out of the cortex region. My LB score dropped by 0.007 which quite surprised me as I was expecting an increase. \n\nHow is your experience with such approach?\n\nKernel: https://www.kaggle.com/keremt/hubmap-fast-submission\n\n**Update:** When including all anatomical structure LB drops 0.003. This means there are mask annotations outside structure polygon.",
    "1125718": "I think so！My result is the same！",
    "1126036": "Yep, been there. Dome that. :)",
    "1129898": "my understanding is that there should be no glomeruli outside of the cortex. Due to annotation errors, I'm afraid that you are penalized for missing otherwise incorrect labels. This is just a guess. A better approach may be to exclude every other anatomical area from training.",
    "1129905": "There is no absolute boundary between regions in the kidney. Therefore, the AS masks provided by subject matter experts serve as a rough guide.",
    "1130025": "Thanks for the clarification!"
  },
  "source": "meta"
}