{
  "id": 207721,
  "title": "Possible data splitting strategy? ",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/207721",
  "author_name": "Debarshi Chanda",
  "post_date": "2020-12-31T01:59:30.902000",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>This is the image with the maximum number of bounding boxes<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4878232%2F6f66dd4ea5e7b770085d2c02c25118eb%2FMax_boxes.jpg?generation=1609379578241049&amp;alt=media\" alt=\"max_boxes_image\"><br>\nWe can clearly see that many boxes overlap to a huge extent because it is annotated by 3 radiologists<br>\nThere are 18 boxes with IoU over 0.5, I have analyzed this in my EDA notebook <a href=\"https://www.kaggle.com/debarshichanda/vinbigdata-chest-x-ray-eda-with-plotly\" target=\"_blank\">here</a><br>\nWhat could be the possible validation strategy for such a dataset?</p>",
  "messages": [
    {
      "id": 1133164,
      "postDate": "2020-12-31T01:59:30.903Z",
      "content": "<p>This is the image with the maximum number of bounding boxes<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4878232%2F6f66dd4ea5e7b770085d2c02c25118eb%2FMax_boxes.jpg?generation=1609379578241049&amp;alt=media\" alt=\"max_boxes_image\"><br>\nWe can clearly see that many boxes overlap to a huge extent because it is annotated by 3 radiologists<br>\nThere are 18 boxes with IoU over 0.5, I have analyzed this in my EDA notebook <a href=\"https://www.kaggle.com/debarshichanda/vinbigdata-chest-x-ray-eda-with-plotly\" target=\"_blank\">here</a><br>\nWhat could be the possible validation strategy for such a dataset?</p>",
      "rawMarkdown": "This is the image with the maximum number of bounding boxes\n![max_boxes_image](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4878232%2F6f66dd4ea5e7b770085d2c02c25118eb%2FMax_boxes.jpg?generation=1609379578241049&alt=media)\nWe can clearly see that many boxes overlap to a huge extent because it is annotated by 3 radiologists\nThere are 18 boxes with IoU over 0.5, I have analyzed this in my EDA notebook [here](https://www.kaggle.com/debarshichanda/vinbigdata-chest-x-ray-eda-with-plotly)\nWhat could be the possible validation strategy for such a dataset?",
      "votes": 6
    },
    {
      "id": 1134682,
      "postDate": "2021-01-01T13:43:33.707Z",
      "content": "<p>Clearly, you'd never want the same image in the training and test set, so definitely group k-fold of some kind. Possibly multi-label stratified group k-fold?</p>",
      "rawMarkdown": "Clearly, you'd never want the same image in the training and test set, so definitely group k-fold of some kind. Possibly multi-label stratified group k-fold?",
      "votes": 1,
      "replies": [
        {
          "id": 1134699,
          "postDate": "2021-01-01T13:54:06.527Z",
          "content": "<p>Heard about it but never tried<br>\nSounds good</p>",
          "rawMarkdown": "Heard about it but never tried\nSounds good"
        },
        {
          "id": 1134732,
          "postDate": "2021-01-01T14:25:54.573Z",
          "content": "<p>I've used this kaggle dataset for using it: <a href=\"https://www.kaggle.com/yasufuminakama/iterative-stratification\" target=\"_blank\">https://www.kaggle.com/yasufuminakama/iterative-stratification</a></p>",
          "rawMarkdown": "I've used this kaggle dataset for using it: https://www.kaggle.com/yasufuminakama/iterative-stratification",
          "votes": 1
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1134682,
      "author_name": "Björn",
      "author_url": "",
      "post_date": "2021-01-01T13:43:33.707000",
      "content": "<p>Clearly, you'd never want the same image in the training and test set, so definitely group k-fold of some kind. Possibly multi-label stratified group k-fold?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1134699,
          "author_name": "Debarshi Chanda",
          "author_url": "",
          "post_date": "2021-01-01T13:54:06.527000",
          "content": "<p>Heard about it but never tried<br>\nSounds good</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1134732,
          "author_name": "Björn",
          "author_url": "",
          "post_date": "2021-01-01T14:25:54.573000",
          "content": "<p>I've used this kaggle dataset for using it: <a href=\"https://www.kaggle.com/yasufuminakama/iterative-stratification\" target=\"_blank\">https://www.kaggle.com/yasufuminakama/iterative-stratification</a></p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1133164": "This is the image with the maximum number of bounding boxes\n![max_boxes_image](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4878232%2F6f66dd4ea5e7b770085d2c02c25118eb%2FMax_boxes.jpg?generation=1609379578241049&alt=media)\nWe can clearly see that many boxes overlap to a huge extent because it is annotated by 3 radiologists\nThere are 18 boxes with IoU over 0.5, I have analyzed this in my EDA notebook [here](https://www.kaggle.com/debarshichanda/vinbigdata-chest-x-ray-eda-with-plotly)\nWhat could be the possible validation strategy for such a dataset?",
    "1134682": "Clearly, you'd never want the same image in the training and test set, so definitely group k-fold of some kind. Possibly multi-label stratified group k-fold?"
  }
}