{
  "id": 333492,
  "title": "Baseline Submissions",
  "url": "/competitions/hubmap-organ-segmentation/discussion/333492",
  "author_name": "",
  "post_date": "2022-06-26T20:21:15.292993800Z",
  "votes": 31,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Here are some basic submissions I've done so that others can use them as a baseline. </p>\n<p>I have open-sourced this <a href=\"https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission\" target=\"_blank\"><strong>introductory notebook</strong></a> for those looking to run similar benchmarks or if you'd like to understand the submission process better.</p>\n<ul>\n<li><a href=\"https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission\" target=\"_blank\">https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission</a></li>\n</ul>\n<hr>\n<table>\n<thead>\n<tr>\n<th><strong>Submission Description</strong></th>\n<th><strong>Public LB</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>All Zeros</td>\n<td>0.0</td>\n</tr>\n<tr>\n<td><strong>All Ones</strong></td>\n<td><strong>0.35</strong></td>\n</tr>\n<tr>\n<td>Random</td>\n<td>0.29</td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.9)</td>\n<td>0.38</td>\n</tr>\n<tr>\n<td><strong>Circle of Ones (r=0.85)</strong></td>\n<td><strong>0.39</strong></td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.8)</td>\n<td>0.38</td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.75)</td>\n<td>0.38</td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.70)</td>\n<td>0.37</td>\n</tr>\n</tbody>\n</table>\n<p><br></p>\n<hr>\n<p><strong>SIDE NOTE ON ALL ONEs SUBMISSION</strong></p>\n<p>I take this to mean that ~21% of the area of all public test images is mask. We will examine the equation to show why:</p>\n<blockquote>\n  <p><strong><code>DICE_SCORE = (2*OVERLAP_area)/(GT_area+PRED_area)</code></strong></p>\n</blockquote>\n<ul>\n<li>Where PRED_area=IMAGE_area (represent w/ A –&nbsp;Image Area)</li>\n<li>Where GT_area=OVERLAP_area (represent w/ X)</li>\n<li>Where DICE_SCORE=0.35</li>\n</ul>\n<blockquote>\n  <p><strong><code>0.35 = (2*X)/(X+A)</code></strong></p>\n</blockquote>\n<p>Therefore, if we multiply both sides by <strong><code>(X+A)</code></strong>…</p>\n<blockquote>\n  <p><strong><code>0.35X + 0.35A = 2X</code></strong></p>\n</blockquote>\n<p>We then combine <strong><code>X</code></strong>s and divide by the newly determined coefficient of X…</p>\n<blockquote>\n  <p><strong><code>X = 0.35A/1.65 = 0.2121A</code></strong></p>\n</blockquote>\n<hr>\n<p><strong>This means that the Ground Truth mask area for the public test set takes up, on-average, 21.2121% of the area of any respective image.</strong></p>\n<p>If anyone wants to check my understanding feel free!</p>\n<hr>\n<p>I hope this helps :)</p>",
  "messages": [
    {
      "id": "1834325",
      "postDate": "06/26/2022 20:21:15",
      "content": "<p>Here are some basic submissions I've done so that others can use them as a baseline. </p>\n<p>I have open-sourced this <a href=\"https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission\" target=\"_blank\"><strong>introductory notebook</strong></a> for those looking to run similar benchmarks or if you'd like to understand the submission process better.</p>\n<ul>\n<li><a href=\"https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission\" target=\"_blank\">https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission</a></li>\n</ul>\n<hr>\n<table>\n<thead>\n<tr>\n<th><strong>Submission Description</strong></th>\n<th><strong>Public LB</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>All Zeros</td>\n<td>0.0</td>\n</tr>\n<tr>\n<td><strong>All Ones</strong></td>\n<td><strong>0.35</strong></td>\n</tr>\n<tr>\n<td>Random</td>\n<td>0.29</td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.9)</td>\n<td>0.38</td>\n</tr>\n<tr>\n<td><strong>Circle of Ones (r=0.85)</strong></td>\n<td><strong>0.39</strong></td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.8)</td>\n<td>0.38</td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.75)</td>\n<td>0.38</td>\n</tr>\n<tr>\n<td>Circle of Ones (r=0.70)</td>\n<td>0.37</td>\n</tr>\n</tbody>\n</table>\n<p><br></p>\n<hr>\n<p><strong>SIDE NOTE ON ALL ONEs SUBMISSION</strong></p>\n<p>I take this to mean that ~21% of the area of all public test images is mask. We will examine the equation to show why:</p>\n<blockquote>\n  <p><strong><code>DICE_SCORE = (2*OVERLAP_area)/(GT_area+PRED_area)</code></strong></p>\n</blockquote>\n<ul>\n<li>Where PRED_area=IMAGE_area (represent w/ A –&nbsp;Image Area)</li>\n<li>Where GT_area=OVERLAP_area (represent w/ X)</li>\n<li>Where DICE_SCORE=0.35</li>\n</ul>\n<blockquote>\n  <p><strong><code>0.35 = (2*X)/(X+A)</code></strong></p>\n</blockquote>\n<p>Therefore, if we multiply both sides by <strong><code>(X+A)</code></strong>…</p>\n<blockquote>\n  <p><strong><code>0.35X + 0.35A = 2X</code></strong></p>\n</blockquote>\n<p>We then combine <strong><code>X</code></strong>s and divide by the newly determined coefficient of X…</p>\n<blockquote>\n  <p><strong><code>X = 0.35A/1.65 = 0.2121A</code></strong></p>\n</blockquote>\n<hr>\n<p><strong>This means that the Ground Truth mask area for the public test set takes up, on-average, 21.2121% of the area of any respective image.</strong></p>\n<p>If anyone wants to check my understanding feel free!</p>\n<hr>\n<p>I hope this helps :)</p>",
      "rawMarkdown": "Here are some basic submissions I've done so that others can use them as a baseline. \n\nI have open-sourced this [**introductory notebook**](https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission) for those looking to run similar benchmarks or if you'd like to understand the submission process better.\n* https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission\n\n---\n\n| **Submission Description** | **Public LB** |\n| --- | --- |\n| All Zeros | 0.0 |\n| **All Ones** | **0.35** |\n| Random | 0.29 |\n| Circle of Ones (r=0.9) | 0.38 |\n| **Circle of Ones (r=0.85)** | **0.39** |\n| Circle of Ones (r=0.8) | 0.38 |\n| Circle of Ones (r=0.75) | 0.38 |\n| Circle of Ones (r=0.70) | 0.37 |\n\n\n<br>\n\n---\n\n**SIDE NOTE ON ALL ONEs SUBMISSION**\n\nI take this to mean that ~21% of the area of all public test images is mask. We will examine the equation to show why:\n\n> **`DICE_SCORE = (2*OVERLAP_area)/(GT_area+PRED_area)`**\n\n* Where PRED_area=IMAGE_area (represent w/ A – Image Area)\n* Where GT_area=OVERLAP_area (represent w/ X)\n* Where DICE_SCORE=0.35\n\n> **`0.35 = (2*X)/(X+A)`**\n\nTherefore, if we multiply both sides by **`(X+A)`**...\n\n> **`0.35X + 0.35A = 2X`**\n\nWe then combine **`X`**s and divide by the newly determined coefficient of X...\n\n> **`X = 0.35A/1.65 = 0.2121A`**\n\n---\n\n**This means that the Ground Truth mask area for the public test set takes up, on-average, 21.2121% of the area of any respective image.**\n\nIf anyone wants to check my understanding feel free!\n\n---\n\nI hope this helps :)",
      "votes": null
    },
    {
      "id": "1834458",
      "postDate": "06/27/2022 00:35:09",
      "content": "<p>there is a trick in this competition (which is a flaw in design):</p>\n<p>\"All images used have at least one FTU\"</p>\n<p>this means that if even if your model predict nothing, it is better to submit something using heuristics.<br>\nthis is made worst, that we know the organ of each test image. so we can \"easily guess\" the potential FTU candidates and obtain some min dice score.<br>\n(e.g. alveolus covers white background region )</p>",
      "rawMarkdown": "there is a trick in this competition (which is a flaw in design):\n\n\"All images used have at least one FTU\"\n\nthis means that if even if your model predict nothing, it is better to submit something using heuristics.\nthis is made worst, that we know the organ of each test image. so we can \"easily guess\" the potential FTU candidates and obtain some min dice score.\n(e.g. alveolus covers white background region )",
      "votes": null
    },
    {
      "id": "1835718",
      "postDate": "06/28/2022 03:08:09",
      "content": "<p>thank, at least giving some distribution info.</p>",
      "rawMarkdown": "thank, at least giving some distribution info.",
      "votes": null
    },
    {
      "id": "1837552",
      "postDate": "06/29/2022 16:59:53",
      "content": "<p>you should conduct the same experiments on your train set.<br>\nthen note the values for train set (HPA images)<br>\nand note the differences compared to public test (HPA +Hubmap images)</p>",
      "rawMarkdown": "you should conduct the same experiments on your train set.\nthen note the values for train set (HPA images)\nand note the differences compared to public test (HPA +Hubmap images)",
      "votes": null
    },
    {
      "id": "1843515",
      "postDate": "07/05/2022 01:03:13",
      "content": "<p><img src=\"https://i.ibb.co/4JYYp4B/Selection-949.png\" alt=\"\"></p>",
      "rawMarkdown": "![](https://i.ibb.co/4JYYp4B/Selection-949.png)",
      "votes": null
    },
    {
      "id": "1885258",
      "postDate": "08/05/2022 03:44:03",
      "content": "<p>Interesting thoughts! My first valid submission is better than random, worse than circle with 0.85 radius. 😂</p>",
      "rawMarkdown": "Interesting thoughts! My first valid submission is better than random, worse than circle with 0.85 radius. 😂",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1834458,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/27/2022 00:35:09",
      "content": "<p>there is a trick in this competition (which is a flaw in design):</p>\n<p>\"All images used have at least one FTU\"</p>\n<p>this means that if even if your model predict nothing, it is better to submit something using heuristics.<br>\nthis is made worst, that we know the organ of each test image. so we can \"easily guess\" the potential FTU candidates and obtain some min dice score.<br>\n(e.g. alveolus covers white background region )</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1835718,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "06/28/2022 03:08:09",
      "content": "<p>thank, at least giving some distribution info.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1837552,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/29/2022 16:59:53",
      "content": "<p>you should conduct the same experiments on your train set.<br>\nthen note the values for train set (HPA images)<br>\nand note the differences compared to public test (HPA +Hubmap images)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1843515,
      "author_name": "thedevastator",
      "author_url": "",
      "post_date": "07/05/2022 01:03:13",
      "content": "<p><img src=\"https://i.ibb.co/4JYYp4B/Selection-949.png\" alt=\"\"></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1885258,
      "author_name": "electro",
      "author_url": "",
      "post_date": "08/05/2022 03:44:03",
      "content": "<p>Interesting thoughts! My first valid submission is better than random, worse than circle with 0.85 radius. 😂</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1834325": "Here are some basic submissions I've done so that others can use them as a baseline. \n\nI have open-sourced this [**introductory notebook**](https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission) for those looking to run similar benchmarks or if you'd like to understand the submission process better.\n* https://www.kaggle.com/code/dschettler8845/hubmap-hpa-organ-segmentation-basic-submission\n\n---\n\n| **Submission Description** | **Public LB** |\n| --- | --- |\n| All Zeros | 0.0 |\n| **All Ones** | **0.35** |\n| Random | 0.29 |\n| Circle of Ones (r=0.9) | 0.38 |\n| **Circle of Ones (r=0.85)** | **0.39** |\n| Circle of Ones (r=0.8) | 0.38 |\n| Circle of Ones (r=0.75) | 0.38 |\n| Circle of Ones (r=0.70) | 0.37 |\n\n\n<br>\n\n---\n\n**SIDE NOTE ON ALL ONEs SUBMISSION**\n\nI take this to mean that ~21% of the area of all public test images is mask. We will examine the equation to show why:\n\n> **`DICE_SCORE = (2*OVERLAP_area)/(GT_area+PRED_area)`**\n\n* Where PRED_area=IMAGE_area (represent w/ A – Image Area)\n* Where GT_area=OVERLAP_area (represent w/ X)\n* Where DICE_SCORE=0.35\n\n> **`0.35 = (2*X)/(X+A)`**\n\nTherefore, if we multiply both sides by **`(X+A)`**...\n\n> **`0.35X + 0.35A = 2X`**\n\nWe then combine **`X`**s and divide by the newly determined coefficient of X...\n\n> **`X = 0.35A/1.65 = 0.2121A`**\n\n---\n\n**This means that the Ground Truth mask area for the public test set takes up, on-average, 21.2121% of the area of any respective image.**\n\nIf anyone wants to check my understanding feel free!\n\n---\n\nI hope this helps :)",
    "1834458": "there is a trick in this competition (which is a flaw in design):\n\n\"All images used have at least one FTU\"\n\nthis means that if even if your model predict nothing, it is better to submit something using heuristics.\nthis is made worst, that we know the organ of each test image. so we can \"easily guess\" the potential FTU candidates and obtain some min dice score.\n(e.g. alveolus covers white background region )",
    "1835718": "thank, at least giving some distribution info.",
    "1837552": "you should conduct the same experiments on your train set.\nthen note the values for train set (HPA images)\nand note the differences compared to public test (HPA +Hubmap images)",
    "1843515": "![](https://i.ibb.co/4JYYp4B/Selection-949.png)",
    "1885258": "Interesting thoughts! My first valid submission is better than random, worse than circle with 0.85 radius. 😂"
  },
  "source": "meta"
}