{
  "id": 114808,
  "title": "Is there any reason that u-net is preferred than other segmentation model?",
  "url": "/competitions/understanding_cloud_organization/discussion/114808",
  "author_name": "",
  "post_date": "2019-10-29T12:07:57.477583600Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I'm late joiner of this competition.\nI read few of discussion or kernel.\nAnd I noticed that many user use u-net than other segmentation models.\nDo you have any special reason using it? I think there are many new models.</p>",
  "messages": [
    {
      "id": "660640",
      "postDate": "10/29/2019 12:07:57",
      "content": "<p>I'm late joiner of this competition.\nI read few of discussion or kernel.\nAnd I noticed that many user use u-net than other segmentation models.\nDo you have any special reason using it? I think there are many new models.</p>",
      "rawMarkdown": "I'm late joiner of this competition.\nI read few of discussion or kernel.\nAnd I noticed that many user use u-net than other segmentation models.\nDo you have any special reason using it? I think there are many new models.",
      "votes": null
    },
    {
      "id": "660684",
      "postDate": "10/29/2019 13:19:59",
      "content": "<p>hahaha . Unet is preferable choice because it always gives you a good baseline performance and most of the datasets , it works .  The last Sevestral Competition FPN worked better than UNet , so , I kind of left the Unet halfway through and started experimenting with FPN .  Some people used PSPnet , DeeplabV3+ in the ensemble as well . \nSame can be said about the encoders , preferable choice is always resnet18, resnet34 to start with . Then we can go ahead and see if we can implement efficientnet b0-b7, se-resnet, resnext , seresnext , inception, densenet and other heavier and deep models , sometimes they do better , sometimes they do worse . </p>\n\n<p>Hope I could answer you ,</p>",
      "rawMarkdown": "hahaha . Unet is preferable choice because it always gives you a good baseline performance and most of the datasets , it works .  The last Sevestral Competition FPN worked better than UNet , so , I kind of left the Unet halfway through and started experimenting with FPN .  Some people used PSPnet , DeeplabV3+ in the ensemble as well . \nSame can be said about the encoders , preferable choice is always resnet18, resnet34 to start with . Then we can go ahead and see if we can implement efficientnet b0-b7, se-resnet, resnext , seresnext , inception, densenet and other heavier and deep models , sometimes they do better , sometimes they do worse . \n\nHope I could answer you ,",
      "votes": null
    },
    {
      "id": "662492",
      "postDate": "10/31/2019 16:06:31",
      "content": "<p>Oh, I see :) thanks for your reply!</p>",
      "rawMarkdown": "Oh, I see :) thanks for your reply!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 660684,
      "author_name": "phoenix9032",
      "author_url": "",
      "post_date": "10/29/2019 13:19:59",
      "content": "<p>hahaha . Unet is preferable choice because it always gives you a good baseline performance and most of the datasets , it works .  The last Sevestral Competition FPN worked better than UNet , so , I kind of left the Unet halfway through and started experimenting with FPN .  Some people used PSPnet , DeeplabV3+ in the ensemble as well . \nSame can be said about the encoders , preferable choice is always resnet18, resnet34 to start with . Then we can go ahead and see if we can implement efficientnet b0-b7, se-resnet, resnext , seresnext , inception, densenet and other heavier and deep models , sometimes they do better , sometimes they do worse . </p>\n\n<p>Hope I could answer you ,</p>",
      "votes": null,
      "replies": [
        {
          "id": 662492,
          "author_name": "vanche",
          "author_url": "",
          "post_date": "10/31/2019 16:06:31",
          "content": "<p>Oh, I see :) thanks for your reply!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "660640": "I'm late joiner of this competition.\nI read few of discussion or kernel.\nAnd I noticed that many user use u-net than other segmentation models.\nDo you have any special reason using it? I think there are many new models.",
    "660684": "hahaha . Unet is preferable choice because it always gives you a good baseline performance and most of the datasets , it works .  The last Sevestral Competition FPN worked better than UNet , so , I kind of left the Unet halfway through and started experimenting with FPN .  Some people used PSPnet , DeeplabV3+ in the ensemble as well . \nSame can be said about the encoders , preferable choice is always resnet18, resnet34 to start with . Then we can go ahead and see if we can implement efficientnet b0-b7, se-resnet, resnext , seresnext , inception, densenet and other heavier and deep models , sometimes they do better , sometimes they do worse . \n\nHope I could answer you ,",
    "662492": "Oh, I see :) thanks for your reply!"
  },
  "source": "meta"
}