{
  "id": 630062,
  "title": "Score Improvement: Background Coverage & New Dataset Options",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/630062",
  "author_name": "耶✌",
  "post_date": "2025-11-18T06:05:02.141000",
  "votes": 7,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I believe there are two ways to improve the score next: </p>\n<ol>\n<li>design a model that can effectively identify <strong>background coverage</strong>.</li>\n<li>create an <strong>additional dataset</strong>, because there is a significant difference between the training set images and the test set images, and the number of supplemental images is limited.</li>\n</ol>\n<p>However, I think designing a more robust algorithm for identification is a long-term solution and more in line with the original intention of the competition, but this is quite challenging.\nGood luck!</p>",
  "messages": [
    {
      "id": 3335648,
      "postDate": "2025-11-18T06:05:02.143Z",
      "content": "<p>I believe there are two ways to improve the score next: </p>\n<ol>\n<li>design a model that can effectively identify <strong>background coverage</strong>.</li>\n<li>create an <strong>additional dataset</strong>, because there is a significant difference between the training set images and the test set images, and the number of supplemental images is limited.</li>\n</ol>\n<p>However, I think designing a more robust algorithm for identification is a long-term solution and more in line with the original intention of the competition, but this is quite challenging.\nGood luck!</p>",
      "rawMarkdown": "I believe there are two ways to improve the score next: \n1. design a model that can effectively identify **background coverage**.\n2. create an **additional dataset**, because there is a significant difference between the training set images and the test set images, and the number of supplemental images is limited.\n \nHowever, I think designing a more robust algorithm for identification is a long-term solution and more in line with the original intention of the competition, but this is quite challenging.\nGood luck!",
      "votes": 7
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3335648": "I believe there are two ways to improve the score next: \n1. design a model that can effectively identify **background coverage**.\n2. create an **additional dataset**, because there is a significant difference between the training set images and the test set images, and the number of supplemental images is limited.\n \nHowever, I think designing a more robust algorithm for identification is a long-term solution and more in line with the original intention of the competition, but this is quite challenging.\nGood luck!"
  }
}