{
  "id": 39285,
  "title": "Review about Semantic Segmentation",
  "url": "/competitions/carvana-image-masking-challenge/discussion/39285",
  "author_name": "",
  "post_date": "2017-09-11T10:47:12.162092500Z",
  "votes": 9,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi all,</p>\n\n<p>I have found a useful blog which reviews about Semantic Segmentaion using Deep Learning and it's progress over the years. I hope this will be helpful for everyone who wants to start off this challenge and as well as to the kagglers who wan t to improve scores in this competition. This is the <a href=\"http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review\">link</a></p>\n\n<p><a href=\"http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review\">http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review</a></p>",
  "messages": [
    {
      "id": "220148",
      "postDate": "09/11/2017 10:47:12",
      "content": "<p>Hi all,</p>\n\n<p>I have found a useful blog which reviews about Semantic Segmentaion using Deep Learning and it's progress over the years. I hope this will be helpful for everyone who wants to start off this challenge and as well as to the kagglers who wan t to improve scores in this competition. This is the <a href=\"http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review\">link</a></p>\n\n<p><a href=\"http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review\">http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review</a></p>",
      "rawMarkdown": "Hi all,\n\nI have found a useful blog which reviews about Semantic Segmentaion using Deep Learning and it's progress over the years. I hope this will be helpful for everyone who wants to start off this challenge and as well as to the kagglers who wan t to improve scores in this competition. This is the [link][1]\n\nhttp://blog.qure.ai/notes/semantic-segmentation-deep-learning-review\n\n\n  [1]: http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review",
      "votes": null
    },
    {
      "id": "220224",
      "postDate": "09/11/2017 14:54:31",
      "content": "<p>Seems like most of semantic segmentation networks don''t work well for carvana. I've tried PSP net, RefineNet, Large Kernel Matters from this post and Tiramisu - but I hadn't got good results with them .. </p>",
      "rawMarkdown": "Seems like most of semantic segmentation networks don''t work well for carvana. I've tried PSP net, RefineNet, Large Kernel Matters from this post and Tiramisu - but I hadn't got good results with them ..",
      "votes": null
    },
    {
      "id": "220280",
      "postDate": "09/11/2017 17:37:04",
      "content": "<p>Same here, and I couldn't figure out why? Can someone share some intuition as to why these top performing segmentation nets work worse than UNet for this problem..</p>",
      "rawMarkdown": "Same here, and I couldn't figure out why? Can someone share some intuition as to why these top performing segmentation nets work worse than UNet for this problem..",
      "votes": null
    },
    {
      "id": "220315",
      "postDate": "09/11/2017 20:09:53",
      "content": "<p>Not sure about Unet, but can answer generally for state-of-art approaches in segmentation. Answer is simple - they do not focus at such a great precision which is required and which is main problem of this competition. Those few numbers after a dot sign.\nState-of-art improves overall, solving other difficult problems. But it is quite far from addressing problem of ~1.0 score.\nThere is approaches that focuses on accurate photo segmentation, mainly post-processing, like  CRF, CRF via RNN, Deep refinement, etc. But even they address real scene, not artificial setup.</p>",
      "rawMarkdown": "Not sure about Unet, but can answer generally for state-of-art approaches in segmentation. Answer is simple - they do not focus at such a great precision which is required and which is main problem of this competition. Those few numbers after a dot sign.\nState-of-art improves overall, solving other difficult problems. But it is quite far from addressing problem of ~1.0 score.\nThere is approaches that focuses on accurate photo segmentation, mainly post-processing, like  CRF, CRF via RNN, Deep refinement, etc. But even they address real scene, not artificial setup.",
      "votes": null
    },
    {
      "id": "220316",
      "postDate": "09/11/2017 20:14:02",
      "content": "<p>Got it. Thanks.</p>",
      "rawMarkdown": "Got it. Thanks.",
      "votes": null
    },
    {
      "id": "220621",
      "postDate": "09/12/2017 17:33:31",
      "content": "<p>seems good </p>",
      "rawMarkdown": "seems good",
      "votes": null
    },
    {
      "id": "220761",
      "postDate": "09/13/2017 01:35:29",
      "content": "<p>I also implemented pspnet, refinenet, and these two networks have very slow convergence, but 'large kernel matters' has very good performance in my experiment</p>",
      "rawMarkdown": "I also implemented pspnet, refinenet, and these two networks have very slow convergence, but 'large kernel matters' has very good performance in my experiment",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 220224,
      "author_name": "heyt0ny",
      "author_url": "",
      "post_date": "09/11/2017 14:54:31",
      "content": "<p>Seems like most of semantic segmentation networks don''t work well for carvana. I've tried PSP net, RefineNet, Large Kernel Matters from this post and Tiramisu - but I hadn't got good results with them .. </p>",
      "votes": null,
      "replies": [
        {
          "id": 220280,
          "author_name": "rrqqmm",
          "author_url": "",
          "post_date": "09/11/2017 17:37:04",
          "content": "<p>Same here, and I couldn't figure out why? Can someone share some intuition as to why these top performing segmentation nets work worse than UNet for this problem..</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 220315,
          "author_name": "",
          "author_url": "",
          "post_date": "09/11/2017 20:09:53",
          "content": "<p>Not sure about Unet, but can answer generally for state-of-art approaches in segmentation. Answer is simple - they do not focus at such a great precision which is required and which is main problem of this competition. Those few numbers after a dot sign.\nState-of-art improves overall, solving other difficult problems. But it is quite far from addressing problem of ~1.0 score.\nThere is approaches that focuses on accurate photo segmentation, mainly post-processing, like  CRF, CRF via RNN, Deep refinement, etc. But even they address real scene, not artificial setup.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 220316,
          "author_name": "rrqqmm",
          "author_url": "",
          "post_date": "09/11/2017 20:14:02",
          "content": "<p>Got it. Thanks.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 220761,
          "author_name": "zhanglj",
          "author_url": "",
          "post_date": "09/13/2017 01:35:29",
          "content": "<p>I also implemented pspnet, refinenet, and these two networks have very slow convergence, but 'large kernel matters' has very good performance in my experiment</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 220621,
      "author_name": "maxpayn25",
      "author_url": "",
      "post_date": "09/12/2017 17:33:31",
      "content": "<p>seems good </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "220148": "Hi all,\n\nI have found a useful blog which reviews about Semantic Segmentaion using Deep Learning and it's progress over the years. I hope this will be helpful for everyone who wants to start off this challenge and as well as to the kagglers who wan t to improve scores in this competition. This is the [link][1]\n\nhttp://blog.qure.ai/notes/semantic-segmentation-deep-learning-review\n\n\n  [1]: http://blog.qure.ai/notes/semantic-segmentation-deep-learning-review",
    "220224": "Seems like most of semantic segmentation networks don''t work well for carvana. I've tried PSP net, RefineNet, Large Kernel Matters from this post and Tiramisu - but I hadn't got good results with them ..",
    "220280": "Same here, and I couldn't figure out why? Can someone share some intuition as to why these top performing segmentation nets work worse than UNet for this problem..",
    "220315": "Not sure about Unet, but can answer generally for state-of-art approaches in segmentation. Answer is simple - they do not focus at such a great precision which is required and which is main problem of this competition. Those few numbers after a dot sign.\nState-of-art improves overall, solving other difficult problems. But it is quite far from addressing problem of ~1.0 score.\nThere is approaches that focuses on accurate photo segmentation, mainly post-processing, like  CRF, CRF via RNN, Deep refinement, etc. But even they address real scene, not artificial setup.",
    "220316": "Got it. Thanks.",
    "220621": "seems good",
    "220761": "I also implemented pspnet, refinenet, and these two networks have very slow convergence, but 'large kernel matters' has very good performance in my experiment"
  },
  "source": "meta"
}