{
  "id": 402998,
  "title": "Information loss from reading tif as 8 bit",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/402998",
  "author_name": "",
  "post_date": "2023-04-20T15:53:36.285484600Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Loading the data with cv2.imread() for example, loads the 16 bit tiff images as 8 bit. I experimented a bit and it does not seem that models trained on the full tiff range perform better than those trained with only 8 bit.<br>\nWhat has your experience been so far?</p>\n<p>Another question would be if you can use the 1.6 TB scans from <a href=\"https://scrollprize.org/data\" target=\"_blank\">https://scrollprize.org/data</a> for training better models. I haven't tried this yet as I have lower effort ideas I want to explore first.</p>",
  "messages": [
    {
      "id": "2228534",
      "postDate": "04/20/2023 15:53:36",
      "content": "<p>Loading the data with cv2.imread() for example, loads the 16 bit tiff images as 8 bit. I experimented a bit and it does not seem that models trained on the full tiff range perform better than those trained with only 8 bit.<br>\nWhat has your experience been so far?</p>\n<p>Another question would be if you can use the 1.6 TB scans from <a href=\"https://scrollprize.org/data\" target=\"_blank\">https://scrollprize.org/data</a> for training better models. I haven't tried this yet as I have lower effort ideas I want to explore first.</p>",
      "rawMarkdown": "Loading the data with cv2.imread() for example, loads the 16 bit tiff images as 8 bit. I experimented a bit and it does not seem that models trained on the full tiff range perform better than those trained with only 8 bit.\nWhat has your experience been so far?\n\nAnother question would be if you can use the 1.6 TB scans from https://scrollprize.org/data for training better models. I haven't tried this yet as I have lower effort ideas I want to explore first.",
      "votes": null
    },
    {
      "id": "2229809",
      "postDate": "04/21/2023 17:46:53",
      "content": "<p>Thanks! That's helpful to know about cv2.imread().  It makes sense that the models would not perform as well, since 8-bit images can only contain 256 grayscale values, as opposed to 65,546 with 16-bit.</p>",
      "rawMarkdown": "Thanks! That's helpful to know about cv2.imread().  It makes sense that the models would not perform as well, since 8-bit images can only contain 256 grayscale values, as opposed to 65,546 with 16-bit.",
      "votes": null
    },
    {
      "id": "2230159",
      "postDate": "04/22/2023 04:30:48",
      "content": "<p>I’ve tried both 8bit and 16bit images and in my experiments there was no difference in models’ performance </p>",
      "rawMarkdown": "I’ve tried both 8bit and 16bit images and in my experiments there was no difference in models’ performance",
      "votes": null
    },
    {
      "id": "2230402",
      "postDate": "04/22/2023 10:40:23",
      "content": "<p>Just to clarify: That this could make a moderate difference was my original assumption, but currently the experiments point to the difference being very minor or non-existent. </p>",
      "rawMarkdown": "Just to clarify: That this could make a moderate difference was my original assumption, but currently the experiments point to the difference being very minor or non-existent.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2229809,
      "author_name": "ajland",
      "author_url": "",
      "post_date": "04/21/2023 17:46:53",
      "content": "<p>Thanks! That's helpful to know about cv2.imread().  It makes sense that the models would not perform as well, since 8-bit images can only contain 256 grayscale values, as opposed to 65,546 with 16-bit.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2230402,
          "author_name": "raki21",
          "author_url": "",
          "post_date": "04/22/2023 10:40:23",
          "content": "<p>Just to clarify: That this could make a moderate difference was my original assumption, but currently the experiments point to the difference being very minor or non-existent. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2230159,
      "author_name": "igorkrashenyi",
      "author_url": "",
      "post_date": "04/22/2023 04:30:48",
      "content": "<p>I’ve tried both 8bit and 16bit images and in my experiments there was no difference in models’ performance </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2228534": "Loading the data with cv2.imread() for example, loads the 16 bit tiff images as 8 bit. I experimented a bit and it does not seem that models trained on the full tiff range perform better than those trained with only 8 bit.\nWhat has your experience been so far?\n\nAnother question would be if you can use the 1.6 TB scans from https://scrollprize.org/data for training better models. I haven't tried this yet as I have lower effort ideas I want to explore first.",
    "2229809": "Thanks! That's helpful to know about cv2.imread().  It makes sense that the models would not perform as well, since 8-bit images can only contain 256 grayscale values, as opposed to 65,546 with 16-bit.",
    "2230159": "I’ve tried both 8bit and 16bit images and in my experiments there was no difference in models’ performance",
    "2230402": "Just to clarify: That this could make a moderate difference was my original assumption, but currently the experiments point to the difference being very minor or non-existent."
  },
  "source": "meta"
}