{
  "id": 306129,
  "title": "What helped me in this competition in terms of the current leaderboard",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/306129",
  "author_name": "",
  "post_date": "2022-02-08T08:51:12.201849400Z",
  "votes": 18,
  "comment_count": 8,
  "views": 0,
  "content": "<p>I have tried a lot with around 240 submissions<br>\n<strong>Very high-resolution inference</strong> (like 6000 ~ 9000 size)<br>\n<strong>Adding a naive IOU tracker</strong> (around 0.009+)<br>\nTwo things worked for me a lot</p>",
  "messages": [
    {
      "id": "1681123",
      "postDate": "02/08/2022 08:51:12",
      "content": "<p>I have tried a lot with around 240 submissions<br>\n<strong>Very high-resolution inference</strong> (like 6000 ~ 9000 size)<br>\n<strong>Adding a naive IOU tracker</strong> (around 0.009+)<br>\nTwo things worked for me a lot</p>",
      "rawMarkdown": "I have tried a lot with around 240 submissions\n**Very high-resolution inference** (like 6000 ~ 9000 size)\n**Adding a naive IOU tracker** (around 0.009+)\nTwo things worked for me a lot",
      "votes": null
    },
    {
      "id": "1682172",
      "postDate": "02/09/2022 01:28:02",
      "content": "<p>great work!</p>",
      "rawMarkdown": "great work!",
      "votes": null
    },
    {
      "id": "1682351",
      "postDate": "02/09/2022 05:41:48",
      "content": "<blockquote>\n  <p>Adding a naive IOU tracker </p>\n</blockquote>\n<p>can I ask if you are using norfair or are you using something else?</p>",
      "rawMarkdown": "> Adding a naive IOU tracker \n\ncan I ask if you are using norfair or are you using something else?",
      "votes": null
    },
    {
      "id": "1682580",
      "postDate": "02/09/2022 08:57:22",
      "content": "<p><a href=\"https://www.kaggle.com/deepkim\" target=\"_blank\">@deepkim</a> I am waiting for finish … priv lb …. and siolution summary … I am almost sure that Very high-resolution inference (like 6000 ~ 9000 size) is overfitting … but could be wrong. If no … I will back to this post and write \"sorry for making a fuss\" 👊💪😄</p>",
      "rawMarkdown": "deepkim I am waiting for finish ... priv lb .... and siolution summary ... I am almost sure that Very high-resolution inference (like 6000 ~ 9000 size) is overfitting ... but could be wrong. If no ... I will back to this post and write \"sorry for making a fuss\" 👊💪😄",
      "votes": null
    },
    {
      "id": "1682626",
      "postDate": "02/09/2022 09:46:33",
      "content": "<p>I mean norfair that was published in the public kernel</p>",
      "rawMarkdown": "I mean norfair that was published in the public kernel",
      "votes": null
    },
    {
      "id": "1682631",
      "postDate": "02/09/2022 09:52:18",
      "content": "<p><a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> but I think most public kernels use the method using high resolution trick and a lot of teams use this high res inference.<br>\nAnyway if priv lb deviates from this phenomenon, most of us will fall simultaneously,  So I think the loss will be very small though a small number of people will be lucky.  <br>\nand Thanks for your great contribution for this competition,, I learned a lot from you</p>",
      "rawMarkdown": "remekkinas but I think most public kernels use the method using high resolution trick and a lot of teams use this high res inference.\nAnyway if priv lb deviates from this phenomenon, most of us will fall simultaneously,  So I think the loss will be very small though a small number of people will be lucky.  \nand Thanks for your great contribution for this competition,, I learned a lot from you",
      "votes": null
    },
    {
      "id": "1683070",
      "postDate": "02/09/2022 14:53:02",
      "content": "<p>using high resolution with CV(3 video split) will lead to very bad F2 score, all the 3 video_id is the same. That's why the high resolution is overfitting the LB. </p>\n<p>Try to find a way to get a high score with both CV and LB is the key to this competition.</p>",
      "rawMarkdown": "using high resolution with CV(3 video split) will lead to very bad F2 score, all the 3 video_id is the same. That's why the high resolution is overfitting the LB. \n\nTry to find a way to get a high score with both CV and LB is the key to this competition.",
      "votes": null
    },
    {
      "id": "1683119",
      "postDate": "02/09/2022 15:31:53",
      "content": "<p><a href=\"https://www.kaggle.com/snaker\" target=\"_blank\">@snaker</a> I agree. It's natural and also ideal that we need to build a model which works for both cv and lb. at the same time I want to ask  whether you don't use high-resolution images when you make a submission since you think that method lowers F2</p>",
      "rawMarkdown": "snaker I agree. It's natural and also ideal that we need to build a model which works for both cv and lb. at the same time I want to ask  whether you don't use high-resolution images when you make a submission since you think that method lowers F2",
      "votes": null
    },
    {
      "id": "1685620",
      "postDate": "02/11/2022 12:40:25",
      "content": "<p>great work!…..</p>",
      "rawMarkdown": "great work!.....",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1682172,
      "author_name": "wonjunpark",
      "author_url": "",
      "post_date": "02/09/2022 01:28:02",
      "content": "<p>great work!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1682351,
      "author_name": "kennyxie",
      "author_url": "",
      "post_date": "02/09/2022 05:41:48",
      "content": "<blockquote>\n  <p>Adding a naive IOU tracker </p>\n</blockquote>\n<p>can I ask if you are using norfair or are you using something else?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1682626,
          "author_name": "deepkim",
          "author_url": "",
          "post_date": "02/09/2022 09:46:33",
          "content": "<p>I mean norfair that was published in the public kernel</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1682580,
      "author_name": "remekkinas",
      "author_url": "",
      "post_date": "02/09/2022 08:57:22",
      "content": "<p><a href=\"https://www.kaggle.com/deepkim\" target=\"_blank\">@deepkim</a> I am waiting for finish … priv lb …. and siolution summary … I am almost sure that Very high-resolution inference (like 6000 ~ 9000 size) is overfitting … but could be wrong. If no … I will back to this post and write \"sorry for making a fuss\" 👊💪😄</p>",
      "votes": null,
      "replies": [
        {
          "id": 1682631,
          "author_name": "deepkim",
          "author_url": "",
          "post_date": "02/09/2022 09:52:18",
          "content": "<p><a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> but I think most public kernels use the method using high resolution trick and a lot of teams use this high res inference.<br>\nAnyway if priv lb deviates from this phenomenon, most of us will fall simultaneously,  So I think the loss will be very small though a small number of people will be lucky.  <br>\nand Thanks for your great contribution for this competition,, I learned a lot from you</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1683070,
          "author_name": "snaker",
          "author_url": "",
          "post_date": "02/09/2022 14:53:02",
          "content": "<p>using high resolution with CV(3 video split) will lead to very bad F2 score, all the 3 video_id is the same. That's why the high resolution is overfitting the LB. </p>\n<p>Try to find a way to get a high score with both CV and LB is the key to this competition.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1683119,
          "author_name": "deepkim",
          "author_url": "",
          "post_date": "02/09/2022 15:31:53",
          "content": "<p><a href=\"https://www.kaggle.com/snaker\" target=\"_blank\">@snaker</a> I agree. It's natural and also ideal that we need to build a model which works for both cv and lb. at the same time I want to ask  whether you don't use high-resolution images when you make a submission since you think that method lowers F2</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1685620,
      "author_name": "abdulmanankhalid",
      "author_url": "",
      "post_date": "02/11/2022 12:40:25",
      "content": "<p>great work!…..</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1681123": "I have tried a lot with around 240 submissions\n**Very high-resolution inference** (like 6000 ~ 9000 size)\n**Adding a naive IOU tracker** (around 0.009+)\nTwo things worked for me a lot",
    "1682172": "great work!",
    "1682351": "> Adding a naive IOU tracker \n\ncan I ask if you are using norfair or are you using something else?",
    "1682580": "deepkim I am waiting for finish ... priv lb .... and siolution summary ... I am almost sure that Very high-resolution inference (like 6000 ~ 9000 size) is overfitting ... but could be wrong. If no ... I will back to this post and write \"sorry for making a fuss\" 👊💪😄",
    "1682626": "I mean norfair that was published in the public kernel",
    "1682631": "remekkinas but I think most public kernels use the method using high resolution trick and a lot of teams use this high res inference.\nAnyway if priv lb deviates from this phenomenon, most of us will fall simultaneously,  So I think the loss will be very small though a small number of people will be lucky.  \nand Thanks for your great contribution for this competition,, I learned a lot from you",
    "1683070": "using high resolution with CV(3 video split) will lead to very bad F2 score, all the 3 video_id is the same. That's why the high resolution is overfitting the LB. \n\nTry to find a way to get a high score with both CV and LB is the key to this competition.",
    "1683119": "snaker I agree. It's natural and also ideal that we need to build a model which works for both cv and lb. at the same time I want to ask  whether you don't use high-resolution images when you make a submission since you think that method lowers F2",
    "1685620": "great work!....."
  },
  "source": "meta"
}