{
  "id": 39974,
  "title": "LB shake-up possibility!",
  "url": "/competitions/carvana-image-masking-challenge/discussion/39974",
  "author_name": "",
  "post_date": "2017-09-25T10:04:34.389189400Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Given that only 10% of data is used for creating public LB, is there a possibility of huge shakeup? What are the speculations.</p>\n\n<p>Thanks</p>",
  "messages": [
    {
      "id": "224143",
      "postDate": "09/25/2017 10:04:34",
      "content": "<p>Given that only 10% of data is used for creating public LB, is there a possibility of huge shakeup? What are the speculations.</p>\n\n<p>Thanks</p>",
      "rawMarkdown": "Given that only 10% of data is used for creating public LB, is there a possibility of huge shakeup? What are the speculations.\n\nThanks",
      "votes": null
    },
    {
      "id": "224178",
      "postDate": "09/25/2017 12:51:14",
      "content": "<p>I don't think so because the partition has been done randomly. I believe that we will not see big changes in the positions.</p>",
      "rawMarkdown": "I don't think so because the partition has been done randomly. I believe that we will not see big changes in the positions.",
      "votes": null
    },
    {
      "id": "224192",
      "postDate": "09/25/2017 13:47:52",
      "content": "<p>I would agree with ironbar. \nThe split is actually 25/75. \nThis is a pixel based classification and not a picture based classification, which make it unlikely given the resolution of the pictures and the local validation is consistent with the LB.\nI don't think that there will be much shakeup unless competitors used heavy post-processing and try to overfit the LB (ex: Santander competition) using it.</p>",
      "rawMarkdown": "I would agree with ironbar. \nThe split is actually 25/75. \nThis is a pixel based classification and not a picture based classification, which make it unlikely given the resolution of the pictures and the local validation is consistent with the LB.\nI don't think that there will be much shakeup unless competitors used heavy post-processing and try to overfit the LB (ex: Santander competition) using it.",
      "votes": null
    },
    {
      "id": "224330",
      "postDate": "09/25/2017 23:44:37",
      "content": "<p>an average car has about 500,000 pixels. A error of about 55x55 pixels (either missing pixels or false positive pixels) gives dice score of about 0.9970. This is also about a error bounday of 1-pixel thick. So the top kagglers results are very close.</p>\n\n<p>I think human error is also in this range.</p>\n\n<p>It is likely that the ranks on private LB may be different, though the public and private dice scores may be similar (e.g. within +/- 0.0001)</p>\n\n<p>For \"stable\" results, it is important to study the error images to ensure that there are no errors of \"large areas\" for \"some images\". i.e., ensure that errors across all images are smiliar.</p>",
      "rawMarkdown": "an average car has about 500,000 pixels. A error of about 55x55 pixels (either missing pixels or false positive pixels) gives dice score of about 0.9970. This is also about a error bounday of 1-pixel thick. So the top kagglers results are very close.\n\nI think human error is also in this range.\n\nIt is likely that the ranks on private LB may be different, though the public and private dice scores may be similar (e.g. within +/- 0.0001)\n\nFor \"stable\" results, it is important to study the error images to ensure that there are no errors of \"large areas\" for \"some images\". i.e., ensure that errors across all images are smiliar.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 224178,
      "author_name": "ironbar",
      "author_url": "",
      "post_date": "09/25/2017 12:51:14",
      "content": "<p>I don't think so because the partition has been done randomly. I believe that we will not see big changes in the positions.</p>",
      "votes": null,
      "replies": [
        {
          "id": 224192,
          "author_name": "chabir",
          "author_url": "",
          "post_date": "09/25/2017 13:47:52",
          "content": "<p>I would agree with ironbar. \nThe split is actually 25/75. \nThis is a pixel based classification and not a picture based classification, which make it unlikely given the resolution of the pictures and the local validation is consistent with the LB.\nI don't think that there will be much shakeup unless competitors used heavy post-processing and try to overfit the LB (ex: Santander competition) using it.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 224330,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "09/25/2017 23:44:37",
      "content": "<p>an average car has about 500,000 pixels. A error of about 55x55 pixels (either missing pixels or false positive pixels) gives dice score of about 0.9970. This is also about a error bounday of 1-pixel thick. So the top kagglers results are very close.</p>\n\n<p>I think human error is also in this range.</p>\n\n<p>It is likely that the ranks on private LB may be different, though the public and private dice scores may be similar (e.g. within +/- 0.0001)</p>\n\n<p>For \"stable\" results, it is important to study the error images to ensure that there are no errors of \"large areas\" for \"some images\". i.e., ensure that errors across all images are smiliar.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "224143": "Given that only 10% of data is used for creating public LB, is there a possibility of huge shakeup? What are the speculations.\n\nThanks",
    "224178": "I don't think so because the partition has been done randomly. I believe that we will not see big changes in the positions.",
    "224192": "I would agree with ironbar. \nThe split is actually 25/75. \nThis is a pixel based classification and not a picture based classification, which make it unlikely given the resolution of the pictures and the local validation is consistent with the LB.\nI don't think that there will be much shakeup unless competitors used heavy post-processing and try to overfit the LB (ex: Santander competition) using it.",
    "224330": "an average car has about 500,000 pixels. A error of about 55x55 pixels (either missing pixels or false positive pixels) gives dice score of about 0.9970. This is also about a error bounday of 1-pixel thick. So the top kagglers results are very close.\n\nI think human error is also in this range.\n\nIt is likely that the ranks on private LB may be different, though the public and private dice scores may be similar (e.g. within +/- 0.0001)\n\nFor \"stable\" results, it is important to study the error images to ensure that there are no errors of \"large areas\" for \"some images\". i.e., ensure that errors across all images are smiliar."
  },
  "source": "meta"
}