{
  "id": 284059,
  "title": "Question: Challenge Metrics vs. (COCO) Object Detection Metric",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/284059",
  "author_name": "michbi",
  "post_date": "2021-10-29T10:42:36.111000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hey everybody,</p>\n<p>I noticed that multiple publications in the cell instance segmentation domain refer to the challenge metric as \"average precision\" and I was wondering if anyone knows its origin or a good motivation to prefer it over the \"standard\" object detection metric (e.g. from COCO)??</p>\n<p>Compared to the typically used mean average precision (mAP) from the natural image processing domain (e.g. Pascal VOC, COCO, LVIS, …) the predictions are not associated with a probability score. As far as I can see it, this procedure discards valuable information during the evaluation =&gt; Effectively looking at a single working point instead of the whole operating range.</p>\n<p>References:<br>\nCell Instance Segmentation Metric</p>\n<ul>\n<li>this challenge</li>\n<li>data science bowl 2018 (<a href=\"https://www.kaggle.com/c/data-science-bowl-2018\" target=\"_blank\">https://www.kaggle.com/c/data-science-bowl-2018</a>)</li>\n<li>Cellpose paper (<a href=\"https://www.nature.com/articles/s41592-020-01018-x\" target=\"_blank\">https://www.nature.com/articles/s41592-020-01018-x</a>)</li>\n</ul>",
  "messages": [
    {
      "id": 1564549,
      "postDate": "2021-10-29T10:42:36.110Z",
      "content": "<p>Hey everybody,</p>\n<p>I noticed that multiple publications in the cell instance segmentation domain refer to the challenge metric as \"average precision\" and I was wondering if anyone knows its origin or a good motivation to prefer it over the \"standard\" object detection metric (e.g. from COCO)??</p>\n<p>Compared to the typically used mean average precision (mAP) from the natural image processing domain (e.g. Pascal VOC, COCO, LVIS, …) the predictions are not associated with a probability score. As far as I can see it, this procedure discards valuable information during the evaluation =&gt; Effectively looking at a single working point instead of the whole operating range.</p>\n<p>References:<br>\nCell Instance Segmentation Metric</p>\n<ul>\n<li>this challenge</li>\n<li>data science bowl 2018 (<a href=\"https://www.kaggle.com/c/data-science-bowl-2018\" target=\"_blank\">https://www.kaggle.com/c/data-science-bowl-2018</a>)</li>\n<li>Cellpose paper (<a href=\"https://www.nature.com/articles/s41592-020-01018-x\" target=\"_blank\">https://www.nature.com/articles/s41592-020-01018-x</a>)</li>\n</ul>",
      "rawMarkdown": "Hey everybody,\n\nI noticed that multiple publications in the cell instance segmentation domain refer to the challenge metric as \"average precision\" and I was wondering if anyone knows its origin or a good motivation to prefer it over the \"standard\" object detection metric (e.g. from COCO)??\n\nCompared to the typically used mean average precision (mAP) from the natural image processing domain (e.g. Pascal VOC, COCO, LVIS, ...) the predictions are not associated with a probability score. As far as I can see it, this procedure discards valuable information during the evaluation => Effectively looking at a single working point instead of the whole operating range.\n\nReferences:\nCell Instance Segmentation Metric\n- this challenge\n- data science bowl 2018 (https://www.kaggle.com/c/data-science-bowl-2018)\n- Cellpose paper (https://www.nature.com/articles/s41592-020-01018-x)\n",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1564549": "Hey everybody,\n\nI noticed that multiple publications in the cell instance segmentation domain refer to the challenge metric as \"average precision\" and I was wondering if anyone knows its origin or a good motivation to prefer it over the \"standard\" object detection metric (e.g. from COCO)??\n\nCompared to the typically used mean average precision (mAP) from the natural image processing domain (e.g. Pascal VOC, COCO, LVIS, ...) the predictions are not associated with a probability score. As far as I can see it, this procedure discards valuable information during the evaluation => Effectively looking at a single working point instead of the whole operating range.\n\nReferences:\nCell Instance Segmentation Metric\n- this challenge\n- data science bowl 2018 (https://www.kaggle.com/c/data-science-bowl-2018)\n- Cellpose paper (https://www.nature.com/articles/s41592-020-01018-x)\n"
  }
}