{
  "id": 281084,
  "title": "Analysis of Overlaps I.",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/281084",
  "author_name": "",
  "post_date": "2021-10-23T15:54:12.762953300Z",
  "votes": 36,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Some additional analysis regarding the issue <a href=\"https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/280250\" target=\"_blank\">Overlaps - ambiguous or not?</a></p>\n<p>I calculated the number of overlaps / number of cells on each image and the <strong>average</strong> overlaping pixels.</p>\n<p>It seems that the amount of average overlaping pixels on the images is rather small:</p>\n<p><img src=\"https://i.postimg.cc/mgHBXdTJ/histogram-of-avg-area.png\" alt=\"histogram of avg area\"></p>\n<p>Divided into cell categories shows that astro's have a higher mean:<br>\n<img src=\"https://i.postimg.cc/PrMzDg3y/histogram-of-avg-area2.png\" alt=\"cell cat histo\"></p>\n<p>If we take a look at the number of overlaps / number of cells on image (with an arbitrary grouping  by 25 pixel area - just to get a grasp how the larger areas are distributed):</p>\n<p><img src=\"https://i.postimg.cc/6pgF8RGX/overlap-vs-cell-1-v2.png\" alt=\"overlap vs cell\"></p>\n<p>Probably the most interesting if we take a look at the distribution with <strong>different cell types</strong>:<br>\n<img src=\"https://i.postimg.cc/52dt8Hzw/overlap-vs-cell-3.png\" alt=\"cell types\"></p>\n<p>Further analysis would be meaningfull to compare the average cell size vs overlaps, thus these determine how much the overlap areas contribute to the IoU!</p>\n<p>Let me know what you think!</p>",
  "messages": [
    {
      "id": "1555127",
      "postDate": "10/23/2021 15:54:12",
      "content": "<p>Some additional analysis regarding the issue <a href=\"https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/280250\" target=\"_blank\">Overlaps - ambiguous or not?</a></p>\n<p>I calculated the number of overlaps / number of cells on each image and the <strong>average</strong> overlaping pixels.</p>\n<p>It seems that the amount of average overlaping pixels on the images is rather small:</p>\n<p><img src=\"https://i.postimg.cc/mgHBXdTJ/histogram-of-avg-area.png\" alt=\"histogram of avg area\"></p>\n<p>Divided into cell categories shows that astro's have a higher mean:<br>\n<img src=\"https://i.postimg.cc/PrMzDg3y/histogram-of-avg-area2.png\" alt=\"cell cat histo\"></p>\n<p>If we take a look at the number of overlaps / number of cells on image (with an arbitrary grouping  by 25 pixel area - just to get a grasp how the larger areas are distributed):</p>\n<p><img src=\"https://i.postimg.cc/6pgF8RGX/overlap-vs-cell-1-v2.png\" alt=\"overlap vs cell\"></p>\n<p>Probably the most interesting if we take a look at the distribution with <strong>different cell types</strong>:<br>\n<img src=\"https://i.postimg.cc/52dt8Hzw/overlap-vs-cell-3.png\" alt=\"cell types\"></p>\n<p>Further analysis would be meaningfull to compare the average cell size vs overlaps, thus these determine how much the overlap areas contribute to the IoU!</p>\n<p>Let me know what you think!</p>",
      "rawMarkdown": "Some additional analysis regarding the issue [Overlaps - ambiguous or not?](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/280250)\n\nI calculated the number of overlaps / number of cells on each image and the **average** overlaping pixels.\n\n It seems that the amount of average overlaping pixels on the images is rather small:\n\n![histogram of avg area](https://i.postimg.cc/mgHBXdTJ/histogram-of-avg-area.png)\n\nDivided into cell categories shows that astro's have a higher mean:\n![cell cat histo](https://i.postimg.cc/PrMzDg3y/histogram-of-avg-area2.png)\n\nIf we take a look at the number of overlaps / number of cells on image (with an arbitrary grouping  by 25 pixel area - just to get a grasp how the larger areas are distributed):\n\n![overlap vs cell](https://i.postimg.cc/6pgF8RGX/overlap-vs-cell-1-v2.png)\n\nProbably the most interesting if we take a look at the distribution with **different cell types**:\n![cell types](https://i.postimg.cc/52dt8Hzw/overlap-vs-cell-3.png)\n\nFurther analysis would be meaningfull to compare the average cell size vs overlaps, thus these determine how much the overlap areas contribute to the IoU!\n\nLet me know what you think!",
      "votes": null
    },
    {
      "id": "1558030",
      "postDate": "10/26/2021 02:21:10",
      "content": "<p>great------------</p>",
      "rawMarkdown": "great------------",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1558030,
      "author_name": "woodalex",
      "author_url": "",
      "post_date": "10/26/2021 02:21:10",
      "content": "<p>great------------</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1555127": "Some additional analysis regarding the issue [Overlaps - ambiguous or not?](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/280250)\n\nI calculated the number of overlaps / number of cells on each image and the **average** overlaping pixels.\n\n It seems that the amount of average overlaping pixels on the images is rather small:\n\n![histogram of avg area](https://i.postimg.cc/mgHBXdTJ/histogram-of-avg-area.png)\n\nDivided into cell categories shows that astro's have a higher mean:\n![cell cat histo](https://i.postimg.cc/PrMzDg3y/histogram-of-avg-area2.png)\n\nIf we take a look at the number of overlaps / number of cells on image (with an arbitrary grouping  by 25 pixel area - just to get a grasp how the larger areas are distributed):\n\n![overlap vs cell](https://i.postimg.cc/6pgF8RGX/overlap-vs-cell-1-v2.png)\n\nProbably the most interesting if we take a look at the distribution with **different cell types**:\n![cell types](https://i.postimg.cc/52dt8Hzw/overlap-vs-cell-3.png)\n\nFurther analysis would be meaningfull to compare the average cell size vs overlaps, thus these determine how much the overlap areas contribute to the IoU!\n\nLet me know what you think!",
    "1558030": "great------------"
  },
  "source": "meta"
}