{
  "id": 304331,
  "title": "About the real-time aspect of the competition",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/304331",
  "author_name": "",
  "post_date": "2022-01-31T20:04:44.193308600Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have noticed that some (if not most) successful models use high image resolutions to improve their F2 score but is it really applicable for this competition or rather to its end goal?</p>\n<p>As stated in the competition overview:</p>\n<blockquote>\n  <p>The goal of this competition is to accurately identify starfish in <strong>real-time</strong> by building an object detection model trained on underwater videos of coral reefs.</p>\n</blockquote>\n<p>And also:</p>\n<blockquote>\n  <p>To scale up video-based surveying systems, Australia’s national science agency, CSIRO has teamed up with Google to develop innovative machine learning technology that can analyse large image datasets accurately, efficiently, and in <strong>near real-time</strong>.</p>\n</blockquote>\n<p>Now, I am not criticizing slow models since I guess the competition organizers could have made sure to restrict the submission notebook times if they required to (If this is even possible in Kaggle). </p>\n<p>What I want is to open a discussion about how these models would help Google and CSIRO achieve their real-time detection goal. Maybe one of you has more experience in this topic and could enlighten us. My guess so far is that they will use the best models to label video data. Then use this to train a proper real-time object detection model.</p>\n<p>If there are any other ideas please share them :)</p>",
  "messages": [
    {
      "id": "1670726",
      "postDate": "01/31/2022 20:04:44",
      "content": "<p>I have noticed that some (if not most) successful models use high image resolutions to improve their F2 score but is it really applicable for this competition or rather to its end goal?</p>\n<p>As stated in the competition overview:</p>\n<blockquote>\n  <p>The goal of this competition is to accurately identify starfish in <strong>real-time</strong> by building an object detection model trained on underwater videos of coral reefs.</p>\n</blockquote>\n<p>And also:</p>\n<blockquote>\n  <p>To scale up video-based surveying systems, Australia’s national science agency, CSIRO has teamed up with Google to develop innovative machine learning technology that can analyse large image datasets accurately, efficiently, and in <strong>near real-time</strong>.</p>\n</blockquote>\n<p>Now, I am not criticizing slow models since I guess the competition organizers could have made sure to restrict the submission notebook times if they required to (If this is even possible in Kaggle). </p>\n<p>What I want is to open a discussion about how these models would help Google and CSIRO achieve their real-time detection goal. Maybe one of you has more experience in this topic and could enlighten us. My guess so far is that they will use the best models to label video data. Then use this to train a proper real-time object detection model.</p>\n<p>If there are any other ideas please share them :)</p>",
      "rawMarkdown": "I have noticed that some (if not most) successful models use high image resolutions to improve their F2 score but is it really applicable for this competition or rather to its end goal?\n\nAs stated in the competition overview:\n> The goal of this competition is to accurately identify starfish in **real-time** by building an object detection model trained on underwater videos of coral reefs.\n\nAnd also:\n\n> To scale up video-based surveying systems, Australia’s national science agency, CSIRO has teamed up with Google to develop innovative machine learning technology that can analyse large image datasets accurately, efficiently, and in **near real-time**.\n\nNow, I am not criticizing slow models since I guess the competition organizers could have made sure to restrict the submission notebook times if they required to (If this is even possible in Kaggle). \n\nWhat I want is to open a discussion about how these models would help Google and CSIRO achieve their real-time detection goal. Maybe one of you has more experience in this topic and could enlighten us. My guess so far is that they will use the best models to label video data. Then use this to train a proper real-time object detection model.\n\nIf there are any other ideas please share them :)",
      "votes": null
    },
    {
      "id": "1670743",
      "postDate": "01/31/2022 20:34:50",
      "content": "<p>Hey, your point about data labeling is cool! </p>\n<p>I would suggest another point to have a low model. I believe that the goal of that competition is not only to build a killer drone but also to manage a population of COTS. So, population management is not also about killing but also about tracking how many COTS are in the specific area.<br>\nI guess organizers will apply our models to offline videos to have a good statistic about COTS spread. </p>",
      "rawMarkdown": "Hey, your point about data labeling is cool! \n\nI would suggest another point to have a low model. I believe that the goal of that competition is not only to build a killer drone but also to manage a population of COTS. So, population management is not also about killing but also about tracking how many COTS are in the specific area.\nI guess organizers will apply our models to offline videos to have a good statistic about COTS spread.",
      "votes": null
    },
    {
      "id": "1670764",
      "postDate": "01/31/2022 20:56:52",
      "content": "<p>I just found this short video about the project (I should have probably watched it before starting the competition):<br>\n<a href=\"url\" target=\"_blank\">https://www.youtube.com/watch?v=UT2noVDFoaA</a><br>\nAnd it appears they are actually trying to control the population as you suggested! Really cool!<br>\nThey still advertise it as a real-time application though so it would be interesting to see what the next steps are after the competition.</p>",
      "rawMarkdown": "I just found this short video about the project (I should have probably watched it before starting the competition):\n[https://www.youtube.com/watch?v=UT2noVDFoaA](url)\nAnd it appears they are actually trying to control the population as you suggested! Really cool!\nThey still advertise it as a real-time application though so it would be interesting to see what the next steps are after the competition.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1670743,
      "author_name": "meowmeowmeowmeowmeow",
      "author_url": "",
      "post_date": "01/31/2022 20:34:50",
      "content": "<p>Hey, your point about data labeling is cool! </p>\n<p>I would suggest another point to have a low model. I believe that the goal of that competition is not only to build a killer drone but also to manage a population of COTS. So, population management is not also about killing but also about tracking how many COTS are in the specific area.<br>\nI guess organizers will apply our models to offline videos to have a good statistic about COTS spread. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1670764,
          "author_name": "deino37",
          "author_url": "",
          "post_date": "01/31/2022 20:56:52",
          "content": "<p>I just found this short video about the project (I should have probably watched it before starting the competition):<br>\n<a href=\"url\" target=\"_blank\">https://www.youtube.com/watch?v=UT2noVDFoaA</a><br>\nAnd it appears they are actually trying to control the population as you suggested! Really cool!<br>\nThey still advertise it as a real-time application though so it would be interesting to see what the next steps are after the competition.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1670726": "I have noticed that some (if not most) successful models use high image resolutions to improve their F2 score but is it really applicable for this competition or rather to its end goal?\n\nAs stated in the competition overview:\n> The goal of this competition is to accurately identify starfish in **real-time** by building an object detection model trained on underwater videos of coral reefs.\n\nAnd also:\n\n> To scale up video-based surveying systems, Australia’s national science agency, CSIRO has teamed up with Google to develop innovative machine learning technology that can analyse large image datasets accurately, efficiently, and in **near real-time**.\n\nNow, I am not criticizing slow models since I guess the competition organizers could have made sure to restrict the submission notebook times if they required to (If this is even possible in Kaggle). \n\nWhat I want is to open a discussion about how these models would help Google and CSIRO achieve their real-time detection goal. Maybe one of you has more experience in this topic and could enlighten us. My guess so far is that they will use the best models to label video data. Then use this to train a proper real-time object detection model.\n\nIf there are any other ideas please share them :)",
    "1670743": "Hey, your point about data labeling is cool! \n\nI would suggest another point to have a low model. I believe that the goal of that competition is not only to build a killer drone but also to manage a population of COTS. So, population management is not also about killing but also about tracking how many COTS are in the specific area.\nI guess organizers will apply our models to offline videos to have a good statistic about COTS spread.",
    "1670764": "I just found this short video about the project (I should have probably watched it before starting the competition):\n[https://www.youtube.com/watch?v=UT2noVDFoaA](url)\nAnd it appears they are actually trying to control the population as you suggested! Really cool!\nThey still advertise it as a real-time application though so it would be interesting to see what the next steps are after the competition."
  },
  "source": "meta"
}