{
  "id": 417314,
  "title": "yet another puzzle game?",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/417314",
  "author_name": "",
  "post_date": "2023-06-15T08:02:23.423537300Z",
  "votes": 18,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Some visualization, I think rebuilding the WSI should be the key for this game, </p>\n<p>The host will update the doc soon, we can access the <code>tile_meta.csv</code> of the test set when evaluation.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F40a29ae10f9da336ecb55992d19dd6d7%2FSCR-20230615-ocpk.jpeg?generation=1686815998402045&amp;alt=media\" alt=\"\"></p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F3cc0181d4f61b2d1b358cdde03c60c65%2FSCR-20230615-oddm.jpeg?generation=1686816084811384&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": "2303286",
      "postDate": "06/15/2023 08:02:23",
      "content": "<p>Some visualization, I think rebuilding the WSI should be the key for this game, </p>\n<p>The host will update the doc soon, we can access the <code>tile_meta.csv</code> of the test set when evaluation.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F40a29ae10f9da336ecb55992d19dd6d7%2FSCR-20230615-ocpk.jpeg?generation=1686815998402045&amp;alt=media\" alt=\"\"></p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F3cc0181d4f61b2d1b358cdde03c60c65%2FSCR-20230615-oddm.jpeg?generation=1686816084811384&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "Some visualization, I think rebuilding the WSI should be the key for this game, ~~but it seems like we do not have  the `tile_meta.csv` for test set.( it'll be very cool to have that though), maybe try some edge detection to solve the puzzle in test set?~~\n\nThe host will update the doc soon, we can access the `tile_meta.csv` of the test set when evaluation.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F40a29ae10f9da336ecb55992d19dd6d7%2FSCR-20230615-ocpk.jpeg?generation=1686815998402045&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F3cc0181d4f61b2d1b358cdde03c60c65%2FSCR-20230615-oddm.jpeg?generation=1686816084811384&alt=media)",
      "votes": null
    },
    {
      "id": "2303294",
      "postDate": "06/15/2023 08:09:59",
      "content": "<p>hmm.. after peeking the whole WSI, maybe semantic segmentation is more suitable for the task than instance segmentation.</p>",
      "rawMarkdown": "hmm.. after peeking the whole WSI, maybe semantic segmentation is more suitable for the task than instance segmentation.",
      "votes": null
    },
    {
      "id": "2303542",
      "postDate": "06/15/2023 10:48:30",
      "content": "<p>Can this task be solved using semantic segmentation?</p>",
      "rawMarkdown": "Can this task be solved using semantic segmentation?",
      "votes": null
    },
    {
      "id": "2303733",
      "postDate": "06/15/2023 12:45:20",
      "content": "<p>I think so. There are no collapse with class 1 mask.</p>",
      "rawMarkdown": "I think so. There are no collapse with class 1 mask.",
      "votes": null
    },
    {
      "id": "2315424",
      "postDate": "06/24/2023 06:03:07",
      "content": "<p>zoom and focus approach.</p>\n<p>you can perform semantic segmentation at larger image (e.g. fuse 2x2 tiles) and lower resolution, etc</p>\n<p>[focus]<br>\nthen generate prompts (potential candiates) from the above results.<br>\ninput each prompt (e.g. random pixel in the above prediction mask)  into e.g. SAM to get instance segmentation<br>\nhere, you can use larger image resolution</p>\n<hr>\n<p>one important tip:</p>\n<ul>\n<li>boundary object are hard to segment. you can improve results by using context from neighbour tile</li>\n</ul>",
      "rawMarkdown": "zoom and focus approach.\n\nyou can perform semantic segmentation at larger image (e.g. fuse 2x2 tiles) and lower resolution, etc\n\n[focus]\nthen generate prompts (potential candiates) from the above results.\ninput each prompt (e.g. random pixel in the above prediction mask)  into e.g. SAM to get instance segmentation\nhere, you can use larger image resolution\n\n---\n\none important tip:\n- boundary object are hard to segment. you can improve results by using context from neighbour tile",
      "votes": null
    },
    {
      "id": "2318381",
      "postDate": "06/26/2023 10:25:15",
      "content": "<p>How much improvement can bring</p>",
      "rawMarkdown": "How much improvement can bring",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2303294,
      "author_name": "snaker",
      "author_url": "",
      "post_date": "06/15/2023 08:09:59",
      "content": "<p>hmm.. after peeking the whole WSI, maybe semantic segmentation is more suitable for the task than instance segmentation.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2303542,
          "author_name": "bent1e",
          "author_url": "",
          "post_date": "06/15/2023 10:48:30",
          "content": "<p>Can this task be solved using semantic segmentation?</p>",
          "votes": null,
          "replies": [
            {
              "id": 2303733,
              "author_name": "snaker",
              "author_url": "",
              "post_date": "06/15/2023 12:45:20",
              "content": "<p>I think so. There are no collapse with class 1 mask.</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2315424,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/24/2023 06:03:07",
      "content": "<p>zoom and focus approach.</p>\n<p>you can perform semantic segmentation at larger image (e.g. fuse 2x2 tiles) and lower resolution, etc</p>\n<p>[focus]<br>\nthen generate prompts (potential candiates) from the above results.<br>\ninput each prompt (e.g. random pixel in the above prediction mask)  into e.g. SAM to get instance segmentation<br>\nhere, you can use larger image resolution</p>\n<hr>\n<p>one important tip:</p>\n<ul>\n<li>boundary object are hard to segment. you can improve results by using context from neighbour tile</li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 2318381,
          "author_name": "bent1e",
          "author_url": "",
          "post_date": "06/26/2023 10:25:15",
          "content": "<p>How much improvement can bring</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2303286": "Some visualization, I think rebuilding the WSI should be the key for this game, ~~but it seems like we do not have  the `tile_meta.csv` for test set.( it'll be very cool to have that though), maybe try some edge detection to solve the puzzle in test set?~~\n\nThe host will update the doc soon, we can access the `tile_meta.csv` of the test set when evaluation.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F40a29ae10f9da336ecb55992d19dd6d7%2FSCR-20230615-ocpk.jpeg?generation=1686815998402045&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F713873%2F3cc0181d4f61b2d1b358cdde03c60c65%2FSCR-20230615-oddm.jpeg?generation=1686816084811384&alt=media)",
    "2303294": "hmm.. after peeking the whole WSI, maybe semantic segmentation is more suitable for the task than instance segmentation.",
    "2303542": "Can this task be solved using semantic segmentation?",
    "2303733": "I think so. There are no collapse with class 1 mask.",
    "2315424": "zoom and focus approach.\n\nyou can perform semantic segmentation at larger image (e.g. fuse 2x2 tiles) and lower resolution, etc\n\n[focus]\nthen generate prompts (potential candiates) from the above results.\ninput each prompt (e.g. random pixel in the above prediction mask)  into e.g. SAM to get instance segmentation\nhere, you can use larger image resolution\n\n---\n\none important tip:\n- boundary object are hard to segment. you can improve results by using context from neighbour tile",
    "2318381": "How much improvement can bring"
  },
  "source": "meta"
}