{
  "id": 416168,
  "title": "Manually refining training data (again)",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/416168",
  "author_name": "Brett Olsen",
  "post_date": "2023-06-10T00:35:25.809000",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>As I discussed in my previous comment (<a href=\"https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/414686)\" target=\"_blank\">https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/414686)</a>, the training data provided by the organizers is less accurate than one might wish. </p>\n<p>This is the second set of manually adjusted training data.  Previously I had adjusted the masks; now I am releasing manually adjusted ink labels for two of the fragments, available here: <a href=\"https://www.kaggle.com/datasets/brettolsen/updated-ink-labels?select=1_adjusted_inklabels.png\" target=\"_blank\">https://www.kaggle.com/datasets/brettolsen/updated-ink-labels?select=1_adjusted_inklabels.png</a></p>\n<p>As before, these should be usable as drop-in replacements for those provided by the organizers.  Please try them out for model training and let me know in the comments if you see any improvements. </p>",
  "messages": [
    {
      "id": 2294310,
      "postDate": "2023-06-10T00:35:25.810Z",
      "content": "<p>As I discussed in my previous comment (<a href=\"https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/414686)\" target=\"_blank\">https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/414686)</a>, the training data provided by the organizers is less accurate than one might wish. </p>\n<p>This is the second set of manually adjusted training data.  Previously I had adjusted the masks; now I am releasing manually adjusted ink labels for two of the fragments, available here: <a href=\"https://www.kaggle.com/datasets/brettolsen/updated-ink-labels?select=1_adjusted_inklabels.png\" target=\"_blank\">https://www.kaggle.com/datasets/brettolsen/updated-ink-labels?select=1_adjusted_inklabels.png</a></p>\n<p>As before, these should be usable as drop-in replacements for those provided by the organizers.  Please try them out for model training and let me know in the comments if you see any improvements. </p>",
      "rawMarkdown": "As I discussed in my previous comment (https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/414686), the training data provided by the organizers is less accurate than one might wish. \n\nThis is the second set of manually adjusted training data.  Previously I had adjusted the masks; now I am releasing manually adjusted ink labels for two of the fragments, available here: https://www.kaggle.com/datasets/brettolsen/updated-ink-labels?select=1_adjusted_inklabels.png\n\nAs before, these should be usable as drop-in replacements for those provided by the organizers.  Please try them out for model training and let me know in the comments if you see any improvements. ",
      "votes": 1
    },
    {
      "id": 2294548,
      "postDate": "2023-06-10T06:34:23.060Z",
      "content": "<p>Interestingly; but what if the test data has similar errors then it might be better for models to learn these error patterns as well?</p>",
      "rawMarkdown": "Interestingly; but what if the test data has similar errors then it might be better for models to learn these error patterns as well?",
      "replies": [
        {
          "id": 2294555,
          "postDate": "2023-06-10T06:42:19.747Z",
          "content": "<p>Possibly!  Try it out and see what the results are.</p>",
          "rawMarkdown": "Possibly!  Try it out and see what the results are.",
          "replies": [
            {
              "id": 2295915,
              "postDate": "2023-06-11T11:43:20.387Z",
              "rawMarkdown": "",
              "isDeleted": true
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2294548,
      "author_name": "Vcol",
      "author_url": "",
      "post_date": "2023-06-10T06:34:23.060000",
      "content": "<p>Interestingly; but what if the test data has similar errors then it might be better for models to learn these error patterns as well?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 2294555,
          "author_name": "Brett Olsen",
          "author_url": "",
          "post_date": "2023-06-10T06:42:19.747000",
          "content": "<p>Possibly!  Try it out and see what the results are.</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2295915,
              "author_name": "",
              "author_url": "",
              "post_date": "2023-06-11T11:43:20.387000",
              "content": "",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2294310": "As I discussed in my previous comment (https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/414686), the training data provided by the organizers is less accurate than one might wish. \n\nThis is the second set of manually adjusted training data.  Previously I had adjusted the masks; now I am releasing manually adjusted ink labels for two of the fragments, available here: https://www.kaggle.com/datasets/brettolsen/updated-ink-labels?select=1_adjusted_inklabels.png\n\nAs before, these should be usable as drop-in replacements for those provided by the organizers.  Please try them out for model training and let me know in the comments if you see any improvements. ",
    "2294548": "Interestingly; but what if the test data has similar errors then it might be better for models to learn these error patterns as well?"
  }
}