{
  "id": 217231,
  "title": "Rely on LB/CV score to tune the model",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/217231",
  "author_name": "Issac",
  "post_date": "2021-02-06T00:14:18.648000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hey guys!</p>\n<p>One thing I've found to be really confusing is tuning the threshold for post-processing (nms/…). Because CV score and LB score are different a lot, which seems like a <a href=\"https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/214284\" target=\"_blank\">common thing for everybody</a>. At first I thought it was because I didn't fuse the annotation bboxes, but even after I did the bbox fusion pre-processing using whatever methods, CV map still differs a lot from LB. So I'm wondering how much the test annotations defer from the given annotations?</p>\n<p>Is the only choice utilizing the 5 chances of submission everyday for tuning this? Or do you guys have any other insights about this problem?</p>",
  "messages": [
    {
      "id": 1188087,
      "postDate": "2021-02-06T00:14:18.647Z",
      "content": "<p>Hey guys!</p>\n<p>One thing I've found to be really confusing is tuning the threshold for post-processing (nms/…). Because CV score and LB score are different a lot, which seems like a <a href=\"https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/214284\" target=\"_blank\">common thing for everybody</a>. At first I thought it was because I didn't fuse the annotation bboxes, but even after I did the bbox fusion pre-processing using whatever methods, CV map still differs a lot from LB. So I'm wondering how much the test annotations defer from the given annotations?</p>\n<p>Is the only choice utilizing the 5 chances of submission everyday for tuning this? Or do you guys have any other insights about this problem?</p>",
      "rawMarkdown": "Hey guys!\n\n\nOne thing I've found to be really confusing is tuning the threshold for post-processing (nms/...). Because CV score and LB score are different a lot, which seems like a [common thing for everybody](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/214284). At first I thought it was because I didn't fuse the annotation bboxes, but even after I did the bbox fusion pre-processing using whatever methods, CV map still differs a lot from LB. So I'm wondering how much the test annotations defer from the given annotations?\n\nIs the only choice utilizing the 5 chances of submission everyday for tuning this? Or do you guys have any other insights about this problem?",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1188087": "Hey guys!\n\n\nOne thing I've found to be really confusing is tuning the threshold for post-processing (nms/...). Because CV score and LB score are different a lot, which seems like a [common thing for everybody](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/214284). At first I thought it was because I didn't fuse the annotation bboxes, but even after I did the bbox fusion pre-processing using whatever methods, CV map still differs a lot from LB. So I'm wondering how much the test annotations defer from the given annotations?\n\nIs the only choice utilizing the 5 chances of submission everyday for tuning this? Or do you guys have any other insights about this problem?"
  }
}