{
  "id": 26641,
  "title": "To deal with small amount of training data with deep learning ",
  "url": "/competitions/dstl-satellite-imagery-feature-detection/discussion/26641",
  "author_name": "",
  "post_date": "2016-12-18T12:31:54.130Z",
  "votes": 8,
  "comment_count": 1,
  "views": 657,
  "content": "<p>Hi,</p>\n\n<p>I think major problem in this competition is a few training dataset compared to test dataset. I found some papers which deal with the problem, so I want to share with you.</p>\n\n<ol>\n<li>Dan C. Ciresan et.al. (2012), Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images, NIPS</li>\n<li>Olaf RonneBerger et.al. (2015), U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv 1505.04597</li>\n</ol>\n\n<p>These papers deal with images segmentation problem in biological research where the number of images tends to be very few compared to the images used in other research areas. \nI think both papers utilising the uniformity of the image to expand the dataset. As you know, we augment dataset by cropping or rotating the images if data were like MNIST, but to preserve the meaning of the image we cannot augment much. But if the image are uniform we can crop and rotate as much as we want, so that we can expand the dataset as much as we like.</p>\n\n<p>It is possible that my understanding is not correct, so please read the papers above if you are interested in the details. </p>\n\n<p>Let's enjoy the competition!  </p>",
  "messages": [
    {
      "id": "151025",
      "postDate": "12/18/2016 12:31:54",
      "content": "<p>Hi,</p>\n\n<p>I think major problem in this competition is a few training dataset compared to test dataset. I found some papers which deal with the problem, so I want to share with you.</p>\n\n<ol>\n<li>Dan C. Ciresan et.al. (2012), Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images, NIPS</li>\n<li>Olaf RonneBerger et.al. (2015), U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv 1505.04597</li>\n</ol>\n\n<p>These papers deal with images segmentation problem in biological research where the number of images tends to be very few compared to the images used in other research areas. \nI think both papers utilising the uniformity of the image to expand the dataset. As you know, we augment dataset by cropping or rotating the images if data were like MNIST, but to preserve the meaning of the image we cannot augment much. But if the image are uniform we can crop and rotate as much as we want, so that we can expand the dataset as much as we like.</p>\n\n<p>It is possible that my understanding is not correct, so please read the papers above if you are interested in the details. </p>\n\n<p>Let's enjoy the competition!  </p>",
      "rawMarkdown": "Hi,\r\n\r\nI think major problem in this competition is a few training dataset compared to test dataset. I found some papers which deal with the problem, so I want to share with you.\r\n\r\n1. Dan C. Ciresan et.al. (2012), Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images, NIPS\r\n2. Olaf RonneBerger et.al. (2015), U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv 1505.04597\r\n\r\nThese papers deal with images segmentation problem in biological research where the number of images tends to be very few compared to the images used in other research areas. \r\nI think both papers utilising the uniformity of the image to expand the dataset. As you know, we augment dataset by cropping or rotating the images if data were like MNIST, but to preserve the meaning of the image we cannot augment much. But if the image are uniform we can crop and rotate as much as we want, so that we can expand the dataset as much as we like.\r\n\r\nIt is possible that my understanding is not correct, so please read the papers above if you are interested in the details. \r\n\r\nLet's enjoy the competition!",
      "votes": null
    },
    {
      "id": "152763",
      "postDate": "12/28/2016 11:38:41",
      "content": "<p>Seems to be interesting articles. Thanks!</p>\n\n<p>If anyone needs direct links:</p>\n\n<ol>\n<li><a href=\"http://papers.nips.cc/paper/4741-deep-neural-networks-segment-neuronal-membranes-in-electron-microscopy-images.pdf\">Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images</a></li>\n<li><a href=\"https://arxiv.org/pdf/1505.04597v1.pdf\">U-Net: Convolutional Networks for BiomedicalImage Segmentation</a></li>\n</ol>",
      "rawMarkdown": "Seems to be interesting articles. Thanks!\r\n\r\nIf anyone needs direct links:\r\n\r\n 1. [Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images][2]\r\n 2. [U-Net: Convolutional Networks for BiomedicalImage Segmentation][1]\r\n\r\n\r\n  [1]: https://arxiv.org/pdf/1505.04597v1.pdf\r\n  [2]: http://papers.nips.cc/paper/4741-deep-neural-networks-segment-neuronal-membranes-in-electron-microscopy-images.pdf",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 152763,
      "author_name": "quantumdamage",
      "author_url": "",
      "post_date": "12/28/2016 11:38:41",
      "content": "<p>Seems to be interesting articles. Thanks!</p>\n\n<p>If anyone needs direct links:</p>\n\n<ol>\n<li><a href=\"http://papers.nips.cc/paper/4741-deep-neural-networks-segment-neuronal-membranes-in-electron-microscopy-images.pdf\">Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images</a></li>\n<li><a href=\"https://arxiv.org/pdf/1505.04597v1.pdf\">U-Net: Convolutional Networks for BiomedicalImage Segmentation</a></li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "151025": "Hi,\r\n\r\nI think major problem in this competition is a few training dataset compared to test dataset. I found some papers which deal with the problem, so I want to share with you.\r\n\r\n1. Dan C. Ciresan et.al. (2012), Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images, NIPS\r\n2. Olaf RonneBerger et.al. (2015), U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv 1505.04597\r\n\r\nThese papers deal with images segmentation problem in biological research where the number of images tends to be very few compared to the images used in other research areas. \r\nI think both papers utilising the uniformity of the image to expand the dataset. As you know, we augment dataset by cropping or rotating the images if data were like MNIST, but to preserve the meaning of the image we cannot augment much. But if the image are uniform we can crop and rotate as much as we want, so that we can expand the dataset as much as we like.\r\n\r\nIt is possible that my understanding is not correct, so please read the papers above if you are interested in the details. \r\n\r\nLet's enjoy the competition!",
    "152763": "Seems to be interesting articles. Thanks!\r\n\r\nIf anyone needs direct links:\r\n\r\n 1. [Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images][2]\r\n 2. [U-Net: Convolutional Networks for BiomedicalImage Segmentation][1]\r\n\r\n\r\n  [1]: https://arxiv.org/pdf/1505.04597v1.pdf\r\n  [2]: http://papers.nips.cc/paper/4741-deep-neural-networks-segment-neuronal-membranes-in-electron-microscopy-images.pdf"
  },
  "source": "meta"
}