{
  "id": 622943,
  "title": "How to read greek on these scrolls? Images have no Ελληνικά γράμματα. Only lines. ",
  "url": "/competitions/vesuvius-challenge-surface-detection/discussion/622943",
  "author_name": "",
  "post_date": "2025-11-15T15:41:43.408769200Z",
  "votes": 12,
  "comment_count": 4,
  "views": 0,
  "content": "<h2>How to read greek on these ancient scrolls lines?</h2>\n<p>Images have no Ελληνικά γράμματα (greek letters). Only lines.</p>\n<p>Till now, every Kaggle Notebook that I read show only lines and masks, sometimes. On the last Vesuvius competiton, we could read greek letters on the papyrus images. Not this time/data.</p>\n<p>Thanks in advance for clarifying this non-coder doubt.</p>\n<p>Με εκτίμηση,</p>\n<p>Μαρίλια Πράτα. </p>",
  "messages": [
    {
      "id": "3327749",
      "postDate": "11/15/2025 15:41:43",
      "content": "<h2>How to read greek on these ancient scrolls lines?</h2>\n<p>Images have no Ελληνικά γράμματα (greek letters). Only lines.</p>\n<p>Till now, every Kaggle Notebook that I read show only lines and masks, sometimes. On the last Vesuvius competiton, we could read greek letters on the papyrus images. Not this time/data.</p>\n<p>Thanks in advance for clarifying this non-coder doubt.</p>\n<p>Με εκτίμηση,</p>\n<p>Μαρίλια Πράτα. </p>",
      "rawMarkdown": "## How to read greek on these ancient scrolls lines? \n\nImages have no Ελληνικά γράμματα (greek letters). Only lines.\n\nTill now, every Kaggle Notebook that I read show only lines and masks, sometimes. On the last Vesuvius competiton, we could read greek letters on the papyrus images. Not this time/data.\n\nThanks in advance for clarifying this non-coder doubt.\n\nΜε εκτίμηση,\n\nΜαρίλια Πράτα.",
      "votes": null
    },
    {
      "id": "3328194",
      "postDate": "11/15/2025 17:35:19",
      "content": "<p>This competition is about surface detection that is the first step of virtual unwrapping the scrolls. Only after virtual unwrapping we can extract the text.</p>",
      "rawMarkdown": "This competition is about surface detection that is the first step of virtual unwrapping the scrolls. Only after virtual unwrapping we can extract the text.",
      "votes": null
    },
    {
      "id": "3328526",
      "postDate": "11/15/2025 20:08:40",
      "content": "<p>It's challenging to match the image and mask for people without backgrounds. Also, just wondering, did you run any baseline model on the dataset?</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1984321%2F4248e5bbf9c1a09377af6cc93dc4d1f3%2FScreenshot%202025-11-16%20020703.png?generation=1763237241062197&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "It's challenging to match the image and mask for people without backgrounds. Also, just wondering, did you run any baseline model on the dataset?\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1984321%2F4248e5bbf9c1a09377af6cc93dc4d1f3%2FScreenshot%202025-11-16%20020703.png?generation=1763237241062197&alt=media)",
      "votes": null
    },
    {
      "id": "3328535",
      "postDate": "11/15/2025 20:19:55",
      "content": "<p>Yes, we have a baseline.</p>\n<p>A papyrus sheet is made of a layer with horizontal fibers and a layer with vertical fibers. We are interested in the one with horizontal fibers, on which the text is written. Sometimes these two layers are separated, and to someone without previous knowledge the labels could seem inaccurate, but they actually aren't. The segmentation does not necessarily have to be single-voxel accurate, this is why we are using a SurfaceDice with a 2 voxel tolerance in the metrics.</p>",
      "rawMarkdown": "Yes, we have a baseline.\n\nA papyrus sheet is made of a layer with horizontal fibers and a layer with vertical fibers. We are interested in the one with horizontal fibers, on which the text is written. Sometimes these two layers are separated, and to someone without previous knowledge the labels could seem inaccurate, but they actually aren't. The segmentation does not necessarily have to be single-voxel accurate, this is why we are using a SurfaceDice with a 2 voxel tolerance in the metrics.",
      "votes": null
    },
    {
      "id": "3328650",
      "postDate": "11/15/2025 23:21:02",
      "content": "<p>Hi Innat and Angelotti,</p>\n<p>I was intrigued cause at the beginning of the competition Overview was mentioned the papyrus fragility (to open):</p>\n<p>\" but that doesn’t mean <strong>we can’t read them. Can your model help?</strong>  We <strong>first need to find the lines</strong>. \"  </p>\n<p>After reading Angelotti's comment below, I got it (I think ).  It's expected to separate the layers (segmentation) to find  <strong>the  Layer with horizontal fibers, on which the text is written</strong>. So reading is another step.</p>\n<p>By the way, I didn't run any baseline. In fact, I didn't even made  the masks. For the record, I have only Kaggle backgrounds, which isn't much to go further : (</p>\n<p>Thank you both for the explanation, I think I got the goal of the competition. I also hope, to read some Kaggle Notebook showing/highlighting that Layer with Horizontal Fibers which is the focus of the competitors.</p>\n<p>Many thanks.</p>",
      "rawMarkdown": "Hi Innat and Angelotti,\n\nI was intrigued cause at the beginning of the competition Overview was mentioned the papyrus fragility (to open):\n\n\" but that doesn’t mean **we can’t read them. Can your model help?**  We **first need to find the lines**. \"  \n\nAfter reading Angelotti's comment below, I got it (I think ).  It's expected to separate the layers (segmentation) to find  **the  Layer with horizontal fibers, on which the text is written**. So reading is another step.\n\nBy the way, I didn't run any baseline. In fact, I didn't even made  the masks. For the record, I have only Kaggle backgrounds, which isn't much to go further : (\n\nThank you both for the explanation, I think I got the goal of the competition. I also hope, to read some Kaggle Notebook showing/highlighting that Layer with Horizontal Fibers which is the focus of the competitors.\n\nMany thanks.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3328194,
      "author_name": "giorgioangelotti",
      "author_url": "",
      "post_date": "11/15/2025 17:35:19",
      "content": "<p>This competition is about surface detection that is the first step of virtual unwrapping the scrolls. Only after virtual unwrapping we can extract the text.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3328526,
          "author_name": "ipythonx",
          "author_url": "",
          "post_date": "11/15/2025 20:08:40",
          "content": "<p>It's challenging to match the image and mask for people without backgrounds. Also, just wondering, did you run any baseline model on the dataset?</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1984321%2F4248e5bbf9c1a09377af6cc93dc4d1f3%2FScreenshot%202025-11-16%20020703.png?generation=1763237241062197&amp;alt=media\" alt=\"\"></p>",
          "votes": null,
          "replies": [
            {
              "id": 3328535,
              "author_name": "giorgioangelotti",
              "author_url": "",
              "post_date": "11/15/2025 20:19:55",
              "content": "<p>Yes, we have a baseline.</p>\n<p>A papyrus sheet is made of a layer with horizontal fibers and a layer with vertical fibers. We are interested in the one with horizontal fibers, on which the text is written. Sometimes these two layers are separated, and to someone without previous knowledge the labels could seem inaccurate, but they actually aren't. The segmentation does not necessarily have to be single-voxel accurate, this is why we are using a SurfaceDice with a 2 voxel tolerance in the metrics.</p>",
              "votes": null,
              "replies": []
            },
            {
              "id": 3328650,
              "author_name": "mpwolke",
              "author_url": "",
              "post_date": "11/15/2025 23:21:02",
              "content": "<p>Hi Innat and Angelotti,</p>\n<p>I was intrigued cause at the beginning of the competition Overview was mentioned the papyrus fragility (to open):</p>\n<p>\" but that doesn’t mean <strong>we can’t read them. Can your model help?</strong>  We <strong>first need to find the lines</strong>. \"  </p>\n<p>After reading Angelotti's comment below, I got it (I think ).  It's expected to separate the layers (segmentation) to find  <strong>the  Layer with horizontal fibers, on which the text is written</strong>. So reading is another step.</p>\n<p>By the way, I didn't run any baseline. In fact, I didn't even made  the masks. For the record, I have only Kaggle backgrounds, which isn't much to go further : (</p>\n<p>Thank you both for the explanation, I think I got the goal of the competition. I also hope, to read some Kaggle Notebook showing/highlighting that Layer with Horizontal Fibers which is the focus of the competitors.</p>\n<p>Many thanks.</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3327749": "## How to read greek on these ancient scrolls lines? \n\nImages have no Ελληνικά γράμματα (greek letters). Only lines.\n\nTill now, every Kaggle Notebook that I read show only lines and masks, sometimes. On the last Vesuvius competiton, we could read greek letters on the papyrus images. Not this time/data.\n\nThanks in advance for clarifying this non-coder doubt.\n\nΜε εκτίμηση,\n\nΜαρίλια Πράτα.",
    "3328194": "This competition is about surface detection that is the first step of virtual unwrapping the scrolls. Only after virtual unwrapping we can extract the text.",
    "3328526": "It's challenging to match the image and mask for people without backgrounds. Also, just wondering, did you run any baseline model on the dataset?\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1984321%2F4248e5bbf9c1a09377af6cc93dc4d1f3%2FScreenshot%202025-11-16%20020703.png?generation=1763237241062197&alt=media)",
    "3328535": "Yes, we have a baseline.\n\nA papyrus sheet is made of a layer with horizontal fibers and a layer with vertical fibers. We are interested in the one with horizontal fibers, on which the text is written. Sometimes these two layers are separated, and to someone without previous knowledge the labels could seem inaccurate, but they actually aren't. The segmentation does not necessarily have to be single-voxel accurate, this is why we are using a SurfaceDice with a 2 voxel tolerance in the metrics.",
    "3328650": "Hi Innat and Angelotti,\n\nI was intrigued cause at the beginning of the competition Overview was mentioned the papyrus fragility (to open):\n\n\" but that doesn’t mean **we can’t read them. Can your model help?**  We **first need to find the lines**. \"  \n\nAfter reading Angelotti's comment below, I got it (I think ).  It's expected to separate the layers (segmentation) to find  **the  Layer with horizontal fibers, on which the text is written**. So reading is another step.\n\nBy the way, I didn't run any baseline. In fact, I didn't even made  the masks. For the record, I have only Kaggle backgrounds, which isn't much to go further : (\n\nThank you both for the explanation, I think I got the goal of the competition. I also hope, to read some Kaggle Notebook showing/highlighting that Layer with Horizontal Fibers which is the focus of the competitors.\n\nMany thanks."
  },
  "source": "meta"
}