{
  "id": 253737,
  "title": "How to include negative samples during training?",
  "url": "/competitions/siim-covid19-detection/discussion/253737",
  "author_name": "",
  "post_date": "2021-07-18T10:52:35.489357900Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>For some images there aren't bounding boxes, still I want to use them during training so as to reduce  false positve, any idea how to implement it?</p>",
  "messages": [
    {
      "id": "1392067",
      "postDate": "07/18/2021 10:52:35",
      "content": "<p>For some images there aren't bounding boxes, still I want to use them during training so as to reduce  false positve, any idea how to implement it?</p>",
      "rawMarkdown": "For some images there aren't bounding boxes, still I want to use them during training so as to reduce  false positve, any idea how to implement it?",
      "votes": null
    },
    {
      "id": "1392632",
      "postDate": "07/18/2021 23:26:09",
      "content": "<p>For YOLOv5, you just include them in your training images folder, yolo will automatically use them as negative images.</p>\n<p>For classification, the images that have a positive study label (typical, indeterminate or atypical) but have no bounding boxes have not been looked at by the annotators, so you decide how you want to deal with them. They are not necessarily negative, but they are not guaranteed to be positive either.</p>",
      "rawMarkdown": "For YOLOv5, you just include them in your training images folder, yolo will automatically use them as negative images.\n\nFor classification, the images that have a positive study label (typical, indeterminate or atypical) but have no bounding boxes have not been looked at by the annotators, so you decide how you want to deal with them. They are not necessarily negative, but they are not guaranteed to be positive either.",
      "votes": null
    },
    {
      "id": "1392889",
      "postDate": "07/19/2021 07:29:57",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/yousof9\" target=\"_blank\">@yousof9</a>! But how about efficientdet and fasterrcnn?</p>",
      "rawMarkdown": "Thanks @yousof9! But how about efficientdet and fasterrcnn?",
      "votes": null
    },
    {
      "id": "1392891",
      "postDate": "07/19/2021 07:32:28",
      "content": "<p>One thing YOLO bother me is that it divides input images into grids and each grid cell will try to detect objects within them. But for chest xray, you have to look at the full picture to draw inference, so is it still advisible to use YOLO?</p>",
      "rawMarkdown": "One thing YOLO bother me is that it divides input images into grids and each grid cell will try to detect objects within them. But for chest xray, you have to look at the full picture to draw inference, so is it still advisible to use YOLO?",
      "votes": null
    },
    {
      "id": "1395424",
      "postDate": "07/21/2021 07:47:17",
      "content": "<p><a href=\"https://www.kaggle.com/logicng\" target=\"_blank\">@logicng</a> looks like effdet can't do it</p>",
      "rawMarkdown": "logicng looks like effdet can't do it",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1392632,
      "author_name": "yousof9",
      "author_url": "",
      "post_date": "07/18/2021 23:26:09",
      "content": "<p>For YOLOv5, you just include them in your training images folder, yolo will automatically use them as negative images.</p>\n<p>For classification, the images that have a positive study label (typical, indeterminate or atypical) but have no bounding boxes have not been looked at by the annotators, so you decide how you want to deal with them. They are not necessarily negative, but they are not guaranteed to be positive either.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1392889,
      "author_name": "logicng",
      "author_url": "",
      "post_date": "07/19/2021 07:29:57",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/yousof9\" target=\"_blank\">@yousof9</a>! But how about efficientdet and fasterrcnn?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1392891,
      "author_name": "logicng",
      "author_url": "",
      "post_date": "07/19/2021 07:32:28",
      "content": "<p>One thing YOLO bother me is that it divides input images into grids and each grid cell will try to detect objects within them. But for chest xray, you have to look at the full picture to draw inference, so is it still advisible to use YOLO?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1395424,
      "author_name": "jaechanlee",
      "author_url": "",
      "post_date": "07/21/2021 07:47:17",
      "content": "<p><a href=\"https://www.kaggle.com/logicng\" target=\"_blank\">@logicng</a> looks like effdet can't do it</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1392067": "For some images there aren't bounding boxes, still I want to use them during training so as to reduce  false positve, any idea how to implement it?",
    "1392632": "For YOLOv5, you just include them in your training images folder, yolo will automatically use them as negative images.\n\nFor classification, the images that have a positive study label (typical, indeterminate or atypical) but have no bounding boxes have not been looked at by the annotators, so you decide how you want to deal with them. They are not necessarily negative, but they are not guaranteed to be positive either.",
    "1392889": "Thanks @yousof9! But how about efficientdet and fasterrcnn?",
    "1392891": "One thing YOLO bother me is that it divides input images into grids and each grid cell will try to detect objects within them. But for chest xray, you have to look at the full picture to draw inference, so is it still advisible to use YOLO?",
    "1395424": "logicng looks like effdet can't do it"
  },
  "source": "meta"
}