{
  "id": 280583,
  "title": "One to many mapping",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/280583",
  "author_name": "Slawek Biel",
  "post_date": "2021-10-21T19:57:05.630000",
  "votes": 14,
  "comment_count": 0,
  "views": 0,
  "content": "<p>With overlaps in target maps there is a curious case where a single prediction object can have IoU greater than 0.5 with more than one target. Here is an example I've found in my validation set where the prediction (in blue) has scores .74 and .79 with two targets.</p>\n<p>The way I'm handling it in the scoring function is to count it as a single true positve and don't count either target as a false negative. I hope that's how the test code behaves too.<br>\n<img src=\"https://raw.githubusercontent.com/slawekslex/random/main/overlap1.png\" alt=\"\"><br>\n<img src=\"https://raw.githubusercontent.com/slawekslex/random/main/overlap2.png\" alt=\"\"></p>\n<p>If you are curious it's image ID 957d8951b270.</p>",
  "messages": [
    {
      "id": 1553018,
      "postDate": "2021-10-21T19:57:05.630Z",
      "content": "<p>With overlaps in target maps there is a curious case where a single prediction object can have IoU greater than 0.5 with more than one target. Here is an example I've found in my validation set where the prediction (in blue) has scores .74 and .79 with two targets.</p>\n<p>The way I'm handling it in the scoring function is to count it as a single true positve and don't count either target as a false negative. I hope that's how the test code behaves too.<br>\n<img src=\"https://raw.githubusercontent.com/slawekslex/random/main/overlap1.png\" alt=\"\"><br>\n<img src=\"https://raw.githubusercontent.com/slawekslex/random/main/overlap2.png\" alt=\"\"></p>\n<p>If you are curious it's image ID 957d8951b270.</p>",
      "rawMarkdown": "With overlaps in target maps there is a curious case where a single prediction object can have IoU greater than 0.5 with more than one target. Here is an example I've found in my validation set where the prediction (in blue) has scores .74 and .79 with two targets.\n\nThe way I'm handling it in the scoring function is to count it as a single true positve and don't count either target as a false negative. I hope that's how the test code behaves too.\n![](https://raw.githubusercontent.com/slawekslex/random/main/overlap1.png)\n![](https://raw.githubusercontent.com/slawekslex/random/main/overlap2.png)\n\nIf you are curious it's image ID 957d8951b270.",
      "votes": 14
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1553018": "With overlaps in target maps there is a curious case where a single prediction object can have IoU greater than 0.5 with more than one target. Here is an example I've found in my validation set where the prediction (in blue) has scores .74 and .79 with two targets.\n\nThe way I'm handling it in the scoring function is to count it as a single true positve and don't count either target as a false negative. I hope that's how the test code behaves too.\n![](https://raw.githubusercontent.com/slawekslex/random/main/overlap1.png)\n![](https://raw.githubusercontent.com/slawekslex/random/main/overlap2.png)\n\nIf you are curious it's image ID 957d8951b270."
  }
}