{
  "id": 301923,
  "title": "How should add the no-bbox data to the validation data?",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/301923",
  "author_name": "",
  "post_date": "2022-01-20T05:30:20.888746Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In order to calculate the accuracy of high resolution models, l’m considering adding no-bbox images to validation data. In this case, should I add all no-bbox images , or should I keep the ratio of some with-bbox  and no-bbox (about 20-80)?</p>\n<p>Guidance would be appreciated.<br>\nThank you.</p>",
  "messages": [
    {
      "id": "1657321",
      "postDate": "01/20/2022 05:30:20",
      "content": "<p>In order to calculate the accuracy of high resolution models, l’m considering adding no-bbox images to validation data. In this case, should I add all no-bbox images , or should I keep the ratio of some with-bbox  and no-bbox (about 20-80)?</p>\n<p>Guidance would be appreciated.<br>\nThank you.</p>",
      "rawMarkdown": "In order to calculate the accuracy of high resolution models, l’m considering adding no-bbox images to validation data. In this case, should I add all no-bbox images , or should I keep the ratio of some with-bbox  and no-bbox (about 20-80)?\n\nGuidance would be appreciated.\nThank you.",
      "votes": null
    },
    {
      "id": "1657845",
      "postDate": "01/20/2022 13:59:25",
      "content": "<p>I have the same question😂.Have you solved it?</p>",
      "rawMarkdown": "I have the same question😂.Have you solved it?",
      "votes": null
    },
    {
      "id": "1657964",
      "postDate": "01/20/2022 15:57:44",
      "content": "<p>I decided to make a video fold and use all the images of the val video to calculate the CV.<br>\nHowever, I don't know how to do this for groupkfold of sequence etc.</p>",
      "rawMarkdown": "I decided to make a video fold and use all the images of the val video to calculate the CV.\nHowever, I don't know how to do this for groupkfold of sequence etc.",
      "votes": null
    },
    {
      "id": "1657985",
      "postDate": "01/20/2022 16:14:43",
      "content": "<p>I mean if you want an accurate calculation, I'd just use all the validation data. If you split by video id, then use the whole video for calculating F2. If you split by sequences then use the whole validation sequence in accuracy calculation. I don't know if it is the best but it is the best \"real\" example you can use (A good mix of bbox and no bbox). </p>\n<p>Does anyone know if the test data have sequences or is it just one long sequence? </p>",
      "rawMarkdown": "I mean if you want an accurate calculation, I'd just use all the validation data. If you split by video id, then use the whole video for calculating F2. If you split by sequences then use the whole validation sequence in accuracy calculation. I don't know if it is the best but it is the best \"real\" example you can use (A good mix of bbox and no bbox). \n\nDoes anyone know if the test data have sequences or is it just one long sequence?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1657845,
      "author_name": "yangranran",
      "author_url": "",
      "post_date": "01/20/2022 13:59:25",
      "content": "<p>I have the same question😂.Have you solved it?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1657964,
          "author_name": "tomp1121",
          "author_url": "",
          "post_date": "01/20/2022 15:57:44",
          "content": "<p>I decided to make a video fold and use all the images of the val video to calculate the CV.<br>\nHowever, I don't know how to do this for groupkfold of sequence etc.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1657985,
      "author_name": "outwrest",
      "author_url": "",
      "post_date": "01/20/2022 16:14:43",
      "content": "<p>I mean if you want an accurate calculation, I'd just use all the validation data. If you split by video id, then use the whole video for calculating F2. If you split by sequences then use the whole validation sequence in accuracy calculation. I don't know if it is the best but it is the best \"real\" example you can use (A good mix of bbox and no bbox). </p>\n<p>Does anyone know if the test data have sequences or is it just one long sequence? </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1657321": "In order to calculate the accuracy of high resolution models, l’m considering adding no-bbox images to validation data. In this case, should I add all no-bbox images , or should I keep the ratio of some with-bbox  and no-bbox (about 20-80)?\n\nGuidance would be appreciated.\nThank you.",
    "1657845": "I have the same question😂.Have you solved it?",
    "1657964": "I decided to make a video fold and use all the images of the val video to calculate the CV.\nHowever, I don't know how to do this for groupkfold of sequence etc.",
    "1657985": "I mean if you want an accurate calculation, I'd just use all the validation data. If you split by video id, then use the whole video for calculating F2. If you split by sequences then use the whole validation sequence in accuracy calculation. I don't know if it is the best but it is the best \"real\" example you can use (A good mix of bbox and no bbox). \n\nDoes anyone know if the test data have sequences or is it just one long sequence?"
  },
  "source": "meta"
}