{
  "id": 64305,
  "title": "Getting started - How to go from NN to bounding boxes?",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/64305",
  "author_name": "",
  "post_date": "2018-08-28T00:25:16.532007700Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I'm more or less used to image segmentation using neural networks that output black/white pixesl. </p>\n\n<p>But how does one gets this result and transform into discrete bounding boxes?\nAnd, at least for me, the most challenging: how to get one, two, three distinct boxes from one image?</p>\n\n<p>Are there different methods other than neural networks for this kind of challenge?</p>",
  "messages": [
    {
      "id": "376706",
      "postDate": "08/28/2018 00:25:16",
      "content": "<p>I'm more or less used to image segmentation using neural networks that output black/white pixesl. </p>\n\n<p>But how does one gets this result and transform into discrete bounding boxes?\nAnd, at least for me, the most challenging: how to get one, two, three distinct boxes from one image?</p>\n\n<p>Are there different methods other than neural networks for this kind of challenge?</p>",
      "rawMarkdown": "I'm more or less used to image segmentation using neural networks that output black/white pixesl. \n\nBut how does one gets this result and transform into discrete bounding boxes?\nAnd, at least for me, the most challenging: how to get one, two, three distinct boxes from one image?\n\nAre there different methods other than neural networks for this kind of challenge?",
      "votes": null
    },
    {
      "id": "376723",
      "postDate": "08/28/2018 01:30:22",
      "content": "<p>You can treat the boxes as 4 output variables. Top left corner and bottom right or x, y, height, width. </p>",
      "rawMarkdown": "You can treat the boxes as 4 output variables. Top left corner and bottom right or x, y, height, width.",
      "votes": null
    },
    {
      "id": "376767",
      "postDate": "08/28/2018 03:21:44",
      "content": "<p>While this does work in theory, it is also theoretically limited to only outputting one box per input. There are several approaches to getting zero or more bounding boxes as output. Some topics/algorithms for you to research:</p>\n\n<ul>\n<li>YOLO (You Only Look Once)</li>\n<li>R-CNN</li>\n<li>Fast R-CNN</li>\n<li>Faster R-CNN (these names are real)</li>\n<li>SSD (Single Shot Detector)</li>\n<li>NASNet</li>\n</ul>\n\n<p>Object detection is a fun research subject, IMO. Good luck!</p>",
      "rawMarkdown": "While this does work in theory, it is also theoretically limited to only outputting one box per input. There are several approaches to getting zero or more bounding boxes as output. Some topics/algorithms for you to research:\n\n - YOLO (You Only Look Once)\n - R-CNN\n - Fast R-CNN\n - Faster R-CNN (these names are real)\n - SSD (Single Shot Detector)\n - NASNet\n\nObject detection is a fun research subject, IMO. Good luck!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 376723,
      "author_name": "anoukstein",
      "author_url": "",
      "post_date": "08/28/2018 01:30:22",
      "content": "<p>You can treat the boxes as 4 output variables. Top left corner and bottom right or x, y, height, width. </p>",
      "votes": null,
      "replies": [
        {
          "id": 376767,
          "author_name": "formigone",
          "author_url": "",
          "post_date": "08/28/2018 03:21:44",
          "content": "<p>While this does work in theory, it is also theoretically limited to only outputting one box per input. There are several approaches to getting zero or more bounding boxes as output. Some topics/algorithms for you to research:</p>\n\n<ul>\n<li>YOLO (You Only Look Once)</li>\n<li>R-CNN</li>\n<li>Fast R-CNN</li>\n<li>Faster R-CNN (these names are real)</li>\n<li>SSD (Single Shot Detector)</li>\n<li>NASNet</li>\n</ul>\n\n<p>Object detection is a fun research subject, IMO. Good luck!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "376706": "I'm more or less used to image segmentation using neural networks that output black/white pixesl. \n\nBut how does one gets this result and transform into discrete bounding boxes?\nAnd, at least for me, the most challenging: how to get one, two, three distinct boxes from one image?\n\nAre there different methods other than neural networks for this kind of challenge?",
    "376723": "You can treat the boxes as 4 output variables. Top left corner and bottom right or x, y, height, width.",
    "376767": "While this does work in theory, it is also theoretically limited to only outputting one box per input. There are several approaches to getting zero or more bounding boxes as output. Some topics/algorithms for you to research:\n\n - YOLO (You Only Look Once)\n - R-CNN\n - Fast R-CNN\n - Faster R-CNN (these names are real)\n - SSD (Single Shot Detector)\n - NASNet\n\nObject detection is a fun research subject, IMO. Good luck!"
  },
  "source": "meta"
}