{
  "id": 294094,
  "title": "just to confirm:  predictions are not allowed to overlap, but are holes allowed?",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/294094",
  "author_name": "",
  "post_date": "2021-12-08T12:31:02.095992700Z",
  "votes": 15,
  "comment_count": 2,
  "views": 0,
  "content": "<p>i want to make sure my understanding is correct:</p>\n<p><a href=\"https://ibb.co/ZfZCSZD\"><img src=\"https://i.ibb.co/mRx70xW/Selection-999-556.png\" alt=\"Selection-999-556\"></a></p>\n<p>so it is possible to represent object in separate disjoint parts in the submission csv?<br>\ni.e. one row in the csv file can contains several disjoint parts</p>\n<p>if this is allowed, then for the overlap region, we could predict 50% of each of the overlapping objects (in the figure above) to minimise loss</p>\n<p>I asked this question, because whether it is allowed or not actually depends on how the evaluation code is written. (e.g. is there any relabelling of connected components in evaluation code)</p>\n<p>while there is no overlapping mask in the test ground truth, but there is overlapping objects in the test images (I think). the organizer may have converted the overlapping masks into non-overlapping ones. e.g.</p>\n<ul>\n<li>dropping overlapping region</li>\n<li>assign the overlapping region to only one object</li>\n<li>dropping overlapping region and any disjoint parts</li>\n</ul>",
  "messages": [
    {
      "id": "1611962",
      "postDate": "12/08/2021 12:31:02",
      "content": "<p>i want to make sure my understanding is correct:</p>\n<p><a href=\"https://ibb.co/ZfZCSZD\"><img src=\"https://i.ibb.co/mRx70xW/Selection-999-556.png\" alt=\"Selection-999-556\"></a></p>\n<p>so it is possible to represent object in separate disjoint parts in the submission csv?<br>\ni.e. one row in the csv file can contains several disjoint parts</p>\n<p>if this is allowed, then for the overlap region, we could predict 50% of each of the overlapping objects (in the figure above) to minimise loss</p>\n<p>I asked this question, because whether it is allowed or not actually depends on how the evaluation code is written. (e.g. is there any relabelling of connected components in evaluation code)</p>\n<p>while there is no overlapping mask in the test ground truth, but there is overlapping objects in the test images (I think). the organizer may have converted the overlapping masks into non-overlapping ones. e.g.</p>\n<ul>\n<li>dropping overlapping region</li>\n<li>assign the overlapping region to only one object</li>\n<li>dropping overlapping region and any disjoint parts</li>\n</ul>",
      "rawMarkdown": "i want to make sure my understanding is correct:\n\n<a href=\"https://ibb.co/ZfZCSZD\"><img src=\"https://i.ibb.co/mRx70xW/Selection-999-556.png\" alt=\"Selection-999-556\" border=\"0\"></a>\n\nso it is possible to represent object in separate disjoint parts in the submission csv?\ni.e. one row in the csv file can contains several disjoint parts\n\nif this is allowed, then for the overlap region, we could predict 50% of each of the overlapping objects (in the figure above) to minimise loss\n\nI asked this question, because whether it is allowed or not actually depends on how the evaluation code is written. (e.g. is there any relabelling of connected components in evaluation code)\n\nwhile there is no overlapping mask in the test ground truth, but there is overlapping objects in the test images (I think). the organizer may have converted the overlapping masks into non-overlapping ones. e.g.\n- dropping overlapping region\n- assign the overlapping region to only one object\n- dropping overlapping region and any disjoint parts",
      "votes": null
    },
    {
      "id": "1612152",
      "postDate": "12/08/2021 15:40:13",
      "content": "<p>Is It row-based? </p>",
      "rawMarkdown": "Is It row-based?",
      "votes": null
    },
    {
      "id": "1616931",
      "postDate": "12/13/2021 18:03:47",
      "content": "<p>My experience so far is is that several disjoint parts in one row will fail, so added filter to check for chopped cells. But I cannot be conclusive in the matter</p>",
      "rawMarkdown": "My experience so far is is that several disjoint parts in one row will fail, so added filter to check for chopped cells. But I cannot be conclusive in the matter",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1612152,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "12/08/2021 15:40:13",
      "content": "<p>Is It row-based? </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1616931,
      "author_name": "mistag",
      "author_url": "",
      "post_date": "12/13/2021 18:03:47",
      "content": "<p>My experience so far is is that several disjoint parts in one row will fail, so added filter to check for chopped cells. But I cannot be conclusive in the matter</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1611962": "i want to make sure my understanding is correct:\n\n<a href=\"https://ibb.co/ZfZCSZD\"><img src=\"https://i.ibb.co/mRx70xW/Selection-999-556.png\" alt=\"Selection-999-556\" border=\"0\"></a>\n\nso it is possible to represent object in separate disjoint parts in the submission csv?\ni.e. one row in the csv file can contains several disjoint parts\n\nif this is allowed, then for the overlap region, we could predict 50% of each of the overlapping objects (in the figure above) to minimise loss\n\nI asked this question, because whether it is allowed or not actually depends on how the evaluation code is written. (e.g. is there any relabelling of connected components in evaluation code)\n\nwhile there is no overlapping mask in the test ground truth, but there is overlapping objects in the test images (I think). the organizer may have converted the overlapping masks into non-overlapping ones. e.g.\n- dropping overlapping region\n- assign the overlapping region to only one object\n- dropping overlapping region and any disjoint parts",
    "1612152": "Is It row-based?",
    "1616931": "My experience so far is is that several disjoint parts in one row will fail, so added filter to check for chopped cells. But I cannot be conclusive in the matter"
  },
  "source": "meta"
}