{
  "id": 280018,
  "title": "Follow-up baseline after Inoichan's topic (Instance segmentation model)",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/280018",
  "author_name": "",
  "post_date": "2021-10-20T02:35:29.019766300Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hello community,</p>\n<p>Following <a href=\"https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790\" target=\"_blank\">this discussion thread</a>, I just published a <a href=\"https://www.kaggle.com/julian3833/sartorius-starter-torch-mask-r-cnn-lb-0-056\" target=\"_blank\">new starter notebook with a Mask R-CNN with pytorch</a>.</p>\n<p>It scores a LB of <code>0.173</code>, getting to the current 10th position. </p>\n<p>The <a href=\"https://www.kaggle.com/julian3833/sartorius-starter-baseline-torch-u-net\" target=\"_blank\">previous U-net model</a>, which I was expecting to enter a steep improvement regime with quick-wins, hit a ceiling at <code>0.03</code>, no matter what changes I performed 🥲, revealing that possibly U-net is not a good fit for this task.</p>\n<p>I hope the community finds this one helpful too.</p>\n<p>Happy kaggling,<br>\ndataista0</p>",
  "messages": [
    {
      "id": "1550774",
      "postDate": "10/20/2021 02:35:29",
      "content": "<p>Hello community,</p>\n<p>Following <a href=\"https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790\" target=\"_blank\">this discussion thread</a>, I just published a <a href=\"https://www.kaggle.com/julian3833/sartorius-starter-torch-mask-r-cnn-lb-0-056\" target=\"_blank\">new starter notebook with a Mask R-CNN with pytorch</a>.</p>\n<p>It scores a LB of <code>0.173</code>, getting to the current 10th position. </p>\n<p>The <a href=\"https://www.kaggle.com/julian3833/sartorius-starter-baseline-torch-u-net\" target=\"_blank\">previous U-net model</a>, which I was expecting to enter a steep improvement regime with quick-wins, hit a ceiling at <code>0.03</code>, no matter what changes I performed 🥲, revealing that possibly U-net is not a good fit for this task.</p>\n<p>I hope the community finds this one helpful too.</p>\n<p>Happy kaggling,<br>\ndataista0</p>",
      "rawMarkdown": "Hello community,\n\nFollowing [this discussion thread](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790), I just published a [new starter notebook with a Mask R-CNN with pytorch](https://www.kaggle.com/julian3833/sartorius-starter-torch-mask-r-cnn-lb-0-056).\n\nIt scores a LB of `0.173`, getting to the current 10th position. \n\n\nThe [previous U-net model](https://www.kaggle.com/julian3833/sartorius-starter-baseline-torch-u-net), which I was expecting to enter a steep improvement regime with quick-wins, hit a ceiling at `0.03`, no matter what changes I performed 🥲, revealing that possibly U-net is not a good fit for this task.\n\nI hope the community finds this one helpful too.\n\nHappy kaggling,\ndataista0",
      "votes": null
    },
    {
      "id": "1551575",
      "postDate": "10/20/2021 17:59:03",
      "content": "<p>I think the U-net model is not well suited to make so many/detailed segmentations on an image</p>",
      "rawMarkdown": "I think the U-net model is not well suited to make so many/detailed segmentations on an image",
      "votes": null
    },
    {
      "id": "1551668",
      "postDate": "10/20/2021 20:00:51",
      "content": "<p>Yes, I totally agree.</p>",
      "rawMarkdown": "Yes, I totally agree.",
      "votes": null
    },
    {
      "id": "1551863",
      "postDate": "10/20/2021 23:11:51",
      "content": "<p>I have very little experienced in segmentation. May I ask why you think <code>U-net</code> like model is not suited for this task ?</p>",
      "rawMarkdown": "I have very little experienced in segmentation. May I ask why you think `U-net` like model is not suited for this task ?",
      "votes": null
    },
    {
      "id": "1552680",
      "postDate": "10/21/2021 15:54:35",
      "content": "<p>Hey <a href=\"https://www.kaggle.com/drhabib\" target=\"_blank\">@drhabib</a>. I have little experience as well, but the points stated <a href=\"https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790\" target=\"_blank\">here</a> are convincing.</p>\n<p>Splitting the unique U-net mask into individuals seems difficult, since the overlapping of the individuals is high in this dataset, in particular for certain types of images. A one-mask output might be adapted, but it seems difficult. Maybe there are techniques for this, but I don't know.</p>\n<p>Mask R-CNN, on the other hand, returns various masks, one per detected individual, resolving that problem before it even happens.</p>",
      "rawMarkdown": "Hey @drhabib. I have little experience as well, but the points stated [here](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790) are convincing.\n\nSplitting the unique U-net mask into individuals seems difficult, since the overlapping of the individuals is high in this dataset, in particular for certain types of images. A one-mask output might be adapted, but it seems difficult. Maybe there are techniques for this, but I don't know.\n\nMask R-CNN, on the other hand, returns various masks, one per detected individual, resolving that problem before it even happens.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1551575,
      "author_name": "tegzes",
      "author_url": "",
      "post_date": "10/20/2021 17:59:03",
      "content": "<p>I think the U-net model is not well suited to make so many/detailed segmentations on an image</p>",
      "votes": null,
      "replies": [
        {
          "id": 1551668,
          "author_name": "julian3833",
          "author_url": "",
          "post_date": "10/20/2021 20:00:51",
          "content": "<p>Yes, I totally agree.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1551863,
          "author_name": "drhabib",
          "author_url": "",
          "post_date": "10/20/2021 23:11:51",
          "content": "<p>I have very little experienced in segmentation. May I ask why you think <code>U-net</code> like model is not suited for this task ?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1552680,
          "author_name": "julian3833",
          "author_url": "",
          "post_date": "10/21/2021 15:54:35",
          "content": "<p>Hey <a href=\"https://www.kaggle.com/drhabib\" target=\"_blank\">@drhabib</a>. I have little experience as well, but the points stated <a href=\"https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790\" target=\"_blank\">here</a> are convincing.</p>\n<p>Splitting the unique U-net mask into individuals seems difficult, since the overlapping of the individuals is high in this dataset, in particular for certain types of images. A one-mask output might be adapted, but it seems difficult. Maybe there are techniques for this, but I don't know.</p>\n<p>Mask R-CNN, on the other hand, returns various masks, one per detected individual, resolving that problem before it even happens.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1550774": "Hello community,\n\nFollowing [this discussion thread](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790), I just published a [new starter notebook with a Mask R-CNN with pytorch](https://www.kaggle.com/julian3833/sartorius-starter-torch-mask-r-cnn-lb-0-056).\n\nIt scores a LB of `0.173`, getting to the current 10th position. \n\n\nThe [previous U-net model](https://www.kaggle.com/julian3833/sartorius-starter-baseline-torch-u-net), which I was expecting to enter a steep improvement regime with quick-wins, hit a ceiling at `0.03`, no matter what changes I performed 🥲, revealing that possibly U-net is not a good fit for this task.\n\nI hope the community finds this one helpful too.\n\nHappy kaggling,\ndataista0",
    "1551575": "I think the U-net model is not well suited to make so many/detailed segmentations on an image",
    "1551668": "Yes, I totally agree.",
    "1551863": "I have very little experienced in segmentation. May I ask why you think `U-net` like model is not suited for this task ?",
    "1552680": "Hey @drhabib. I have little experience as well, but the points stated [here](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279790) are convincing.\n\nSplitting the unique U-net mask into individuals seems difficult, since the overlapping of the individuals is high in this dataset, in particular for certain types of images. A one-mask output might be adapted, but it seems difficult. Maybe there are techniques for this, but I don't know.\n\nMask R-CNN, on the other hand, returns various masks, one per detected individual, resolving that problem before it even happens."
  },
  "source": "meta"
}