{
  "id": 54500,
  "title": "FYI: Capsule Net for Object Segmentation",
  "url": "/competitions/data-science-bowl-2018/discussion/54500",
  "author_name": "",
  "post_date": "2018-04-14T00:47:11.430272600Z",
  "votes": 17,
  "comment_count": 3,
  "views": 0,
  "content": "<p>It compares unet and capsule net. In fact it is a U-shape capsule net.</p>\n\n<p>\"Capsules for Object Segmentation\"- Rodney LaLonde, Ulas Bagci</p>\n\n<p><a href=\"https://arxiv.org/pdf/1804.04241.pdf\">https://arxiv.org/pdf/1804.04241.pdf</a></p>",
  "messages": [
    {
      "id": "313863",
      "postDate": "04/14/2018 00:47:11",
      "content": "<p>It compares unet and capsule net. In fact it is a U-shape capsule net.</p>\n\n<p>\"Capsules for Object Segmentation\"- Rodney LaLonde, Ulas Bagci</p>\n\n<p><a href=\"https://arxiv.org/pdf/1804.04241.pdf\">https://arxiv.org/pdf/1804.04241.pdf</a></p>",
      "rawMarkdown": "It compares unet and capsule net. In fact it is a U-shape capsule net.\n\n\"Capsules for Object Segmentation\"- Rodney LaLonde, Ulas Bagci\n \nhttps://arxiv.org/pdf/1804.04241.pdf",
      "votes": null
    },
    {
      "id": "313965",
      "postDate": "04/14/2018 07:24:18",
      "content": "<p>A pity they don't report on training speed. A similar thing happened with tiramisu: much slower than Unet and improvements are small comparatively. It'll be interested to see how it performs on with small datasets and generalize beyond those.</p>",
      "rawMarkdown": "A pity they don't report on training speed. A similar thing happened with tiramisu: much slower than Unet and improvements are small comparatively. It'll be interested to see how it performs on with small datasets and generalize beyond those.",
      "votes": null
    },
    {
      "id": "314309",
      "postDate": "04/15/2018 07:51:35",
      "content": "<p>CT Scan image results from the paper Capsules for Object Segmentation.</p>",
      "rawMarkdown": "CT Scan image results from the paper Capsules for Object Segmentation.",
      "votes": null
    },
    {
      "id": "314787",
      "postDate": "04/16/2018 09:17:41",
      "content": "<p>The authors responded about training time via twitter:</p>\n\n<p>Training time on a single Titan Xp GPU: U-Net ~ 30 min/10k iters, SegCaps ~ 90 min/10k iters. U-Net converged in ~ 352k iters, SegCaps in ~ 232k iters, averaged over our 4 training splits. Happy to provide any more info!</p>",
      "rawMarkdown": "The authors responded about training time via twitter:\n\nTraining time on a single Titan Xp GPU: U-Net ~ 30 min/10k iters, SegCaps ~ 90 min/10k iters. U-Net converged in ~ 352k iters, SegCaps in ~ 232k iters, averaged over our 4 training splits. Happy to provide any more info!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 313965,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "04/14/2018 07:24:18",
      "content": "<p>A pity they don't report on training speed. A similar thing happened with tiramisu: much slower than Unet and improvements are small comparatively. It'll be interested to see how it performs on with small datasets and generalize beyond those.</p>",
      "votes": null,
      "replies": [
        {
          "id": 314787,
          "author_name": "antorsae",
          "author_url": "",
          "post_date": "04/16/2018 09:17:41",
          "content": "<p>The authors responded about training time via twitter:</p>\n\n<p>Training time on a single Titan Xp GPU: U-Net ~ 30 min/10k iters, SegCaps ~ 90 min/10k iters. U-Net converged in ~ 352k iters, SegCaps in ~ 232k iters, averaged over our 4 training splits. Happy to provide any more info!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 314309,
      "author_name": "iitrsamrat",
      "author_url": "",
      "post_date": "04/15/2018 07:51:35",
      "content": "<p>CT Scan image results from the paper Capsules for Object Segmentation.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "313863": "It compares unet and capsule net. In fact it is a U-shape capsule net.\n\n\"Capsules for Object Segmentation\"- Rodney LaLonde, Ulas Bagci\n \nhttps://arxiv.org/pdf/1804.04241.pdf",
    "313965": "A pity they don't report on training speed. A similar thing happened with tiramisu: much slower than Unet and improvements are small comparatively. It'll be interested to see how it performs on with small datasets and generalize beyond those.",
    "314309": "CT Scan image results from the paper Capsules for Object Segmentation.",
    "314787": "The authors responded about training time via twitter:\n\nTraining time on a single Titan Xp GPU: U-Net ~ 30 min/10k iters, SegCaps ~ 90 min/10k iters. U-Net converged in ~ 352k iters, SegCaps in ~ 232k iters, averaged over our 4 training splits. Happy to provide any more info!"
  },
  "source": "meta"
}