{
  "id": 70772,
  "title": "where rnn is better than cnn",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/70772",
  "author_name": "hengck23",
  "post_date": "2018-11-07T06:15:59.617000",
  "votes": 11,
  "comment_count": 1,
  "views": 0,
  "content": "<p>see attached:</p>\n\n<p>left: stroke without time, right stroke with time</p>\n\n<p>without stroke, we may think that the drawing consists of small overlapping circles. hence it may be grapes, black berry, etc ...</p>\n\n<p>with stroke, we know that the drawing is more likely to be flower</p>\n\n<p>you can think of the time information as\" free segmentation information\", which edges are connected.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10626/0.52_9512884680456237_blackberry_camouflage_blueberry.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10634/0.76_9509195463704973_grapes_blackberry_blueberry.png\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": 416706,
      "postDate": "2018-11-07T06:15:59.617Z",
      "content": "<p>see attached:</p>\n\n<p>left: stroke without time, right stroke with time</p>\n\n<p>without stroke, we may think that the drawing consists of small overlapping circles. hence it may be grapes, black berry, etc ...</p>\n\n<p>with stroke, we know that the drawing is more likely to be flower</p>\n\n<p>you can think of the time information as\" free segmentation information\", which edges are connected.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10626/0.52_9512884680456237_blackberry_camouflage_blueberry.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10634/0.76_9509195463704973_grapes_blackberry_blueberry.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "see attached:\n\nleft: stroke without time, right stroke with time\n\nwithout stroke, we may think that the drawing consists of small overlapping circles. hence it may be grapes, black berry, etc ...\n\nwith stroke, we know that the drawing is more likely to be flower\n\nyou can think of the time information as\" free segmentation information\", which edges are connected.\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10626/0.52_9512884680456237_blackberry_camouflage_blueberry.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10634/0.76_9509195463704973_grapes_blackberry_blueberry.png",
      "votes": 11
    },
    {
      "id": 418112,
      "postDate": "2018-11-09T09:49:27.963Z",
      "content": "<p>some basic baseline:</p>\n\n<ul>\n<li><p>there are about 10% non-recognized train samples</p></li>\n<li><p>out of all  recognized samples, CNN has about 98% accuracy for the easy classes</p></li>\n<li><p>the samples are label recognized and  non-recognized  using RNN (google quickdraw game)</p></li>\n<li><p>this means that RNN has the potential of boosting CNN  results +2%</p></li>\n<li><p>same of the non-recognized train samples are actually correct drawing. And these can be recognized by CNN.\nthis means that CNN can also boost RNN results</p></li>\n</ul>",
      "rawMarkdown": "some basic baseline:\n\n- there are about 10% non-recognized train samples\n\n- out of all  recognized samples, CNN has about 98% accuracy for the easy classes\n\n- the samples are label recognized and  non-recognized  using RNN (google quickdraw game)\n\n- this means that RNN has the potential of boosting CNN  results +2%\n\n- same of the non-recognized train samples are actually correct drawing. And these can be recognized by CNN.\n  this means that CNN can also boost RNN results\n\n",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 418112,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-11-09T09:49:27.963000",
      "content": "<p>some basic baseline:</p>\n\n<ul>\n<li><p>there are about 10% non-recognized train samples</p></li>\n<li><p>out of all  recognized samples, CNN has about 98% accuracy for the easy classes</p></li>\n<li><p>the samples are label recognized and  non-recognized  using RNN (google quickdraw game)</p></li>\n<li><p>this means that RNN has the potential of boosting CNN  results +2%</p></li>\n<li><p>same of the non-recognized train samples are actually correct drawing. And these can be recognized by CNN.\nthis means that CNN can also boost RNN results</p></li>\n</ul>",
      "votes": 2,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "416706": "see attached:\n\nleft: stroke without time, right stroke with time\n\nwithout stroke, we may think that the drawing consists of small overlapping circles. hence it may be grapes, black berry, etc ...\n\nwith stroke, we know that the drawing is more likely to be flower\n\nyou can think of the time information as\" free segmentation information\", which edges are connected.\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10626/0.52_9512884680456237_blackberry_camouflage_blueberry.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/416706/10634/0.76_9509195463704973_grapes_blackberry_blueberry.png",
    "418112": "some basic baseline:\n\n- there are about 10% non-recognized train samples\n\n- out of all  recognized samples, CNN has about 98% accuracy for the easy classes\n\n- the samples are label recognized and  non-recognized  using RNN (google quickdraw game)\n\n- this means that RNN has the potential of boosting CNN  results +2%\n\n- same of the non-recognized train samples are actually correct drawing. And these can be recognized by CNN.\n  this means that CNN can also boost RNN results\n\n"
  }
}