{
  "id": 416400,
  "title": "Try a better loss function for f0.5 scores",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/416400",
  "author_name": "",
  "post_date": "2023-06-11T07:51:14.326843800Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi, I have found a loss function that allows the model to reduce the number of false positives. 👀 <br>\npaper: <a href=\"url\" target=\"_blank\">https://arxiv.org/abs/2106.14917</a><br>\ngithub: <a href=\"url\" target=\"_blank\">https://github.com/PotatoTian/recall-semseg</a><br>\nAfter some of my observations (not quantified), I think the key to improving scores is to reduce false positives. With little time left, I thought I'd post my opinion first, and subsequently I'll retrain my model and update the comments section if I have better results.🤠</p>",
  "messages": [
    {
      "id": "2295702",
      "postDate": "06/11/2023 07:51:14",
      "content": "<p>Hi, I have found a loss function that allows the model to reduce the number of false positives. 👀 <br>\npaper: <a href=\"url\" target=\"_blank\">https://arxiv.org/abs/2106.14917</a><br>\ngithub: <a href=\"url\" target=\"_blank\">https://github.com/PotatoTian/recall-semseg</a><br>\nAfter some of my observations (not quantified), I think the key to improving scores is to reduce false positives. With little time left, I thought I'd post my opinion first, and subsequently I'll retrain my model and update the comments section if I have better results.🤠</p>",
      "rawMarkdown": "Hi, I have found a loss function that allows the model to reduce the number of false positives. 👀 \npaper: [https://arxiv.org/abs/2106.14917](url)\ngithub: [https://github.com/PotatoTian/recall-semseg](url)\nAfter some of my observations (not quantified), I think the key to improving scores is to reduce false positives. With little time left, I thought I'd post my opinion first, and subsequently I'll retrain my model and update the comments section if I have better results.🤠",
      "votes": null
    },
    {
      "id": "2295713",
      "postDate": "06/11/2023 07:58:41",
      "content": "<p>I think using a variety of models might give a better boost to the score.<br>\nFor example, using one 'aggressive' model and two 'conservative' models might have better results. 🤔</p>",
      "rawMarkdown": "I think using a variety of models might give a better boost to the score.\nFor example, using one 'aggressive' model and two 'conservative' models might have better results. 🤔",
      "votes": null
    },
    {
      "id": "2299988",
      "postDate": "06/13/2023 01:36:17",
      "content": "<p>I only discovered your post today. Did the RecallCE have a good effect?</p>",
      "rawMarkdown": "I only discovered your post today. Did the RecallCE have a good effect?",
      "votes": null
    },
    {
      "id": "2300148",
      "postDate": "06/13/2023 04:00:31",
      "content": "<p>This repo can also be a reference: <a href=\"https://github.com/JunMa11/SegLoss\" target=\"_blank\">https://github.com/JunMa11/SegLoss</a></p>",
      "rawMarkdown": "This repo can also be a reference: https://github.com/JunMa11/SegLoss",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2295713,
      "author_name": "jimmyisme1",
      "author_url": "",
      "post_date": "06/11/2023 07:58:41",
      "content": "<p>I think using a variety of models might give a better boost to the score.<br>\nFor example, using one 'aggressive' model and two 'conservative' models might have better results. 🤔</p>",
      "votes": null,
      "replies": [
        {
          "id": 2299988,
          "author_name": "hangtianyu",
          "author_url": "",
          "post_date": "06/13/2023 01:36:17",
          "content": "<p>I only discovered your post today. Did the RecallCE have a good effect?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2300148,
      "author_name": "junxhuang",
      "author_url": "",
      "post_date": "06/13/2023 04:00:31",
      "content": "<p>This repo can also be a reference: <a href=\"https://github.com/JunMa11/SegLoss\" target=\"_blank\">https://github.com/JunMa11/SegLoss</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2295702": "Hi, I have found a loss function that allows the model to reduce the number of false positives. 👀 \npaper: [https://arxiv.org/abs/2106.14917](url)\ngithub: [https://github.com/PotatoTian/recall-semseg](url)\nAfter some of my observations (not quantified), I think the key to improving scores is to reduce false positives. With little time left, I thought I'd post my opinion first, and subsequently I'll retrain my model and update the comments section if I have better results.🤠",
    "2295713": "I think using a variety of models might give a better boost to the score.\nFor example, using one 'aggressive' model and two 'conservative' models might have better results. 🤔",
    "2299988": "I only discovered your post today. Did the RecallCE have a good effect?",
    "2300148": "This repo can also be a reference: https://github.com/JunMa11/SegLoss"
  },
  "source": "meta"
}