{
  "id": 57947,
  "title": "analyzing results of LB 0.4922",
  "url": "/competitions/trackml-particle-identification/discussion/57947",
  "author_name": "",
  "post_date": "2018-05-31T08:14:42.525396200Z",
  "votes": 12,
  "comment_count": 10,
  "views": 0,
  "content": "<p>as shown below</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/336192/9547/Slide1.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/336192/9549/Slide3.png\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": "336192",
      "postDate": "05/31/2018 08:14:42",
      "content": "<p>as shown below</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/336192/9547/Slide1.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/336192/9549/Slide3.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "as shown below\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/336192/9547/Slide1.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/336192/9549/Slide3.png",
      "votes": null
    },
    {
      "id": "336376",
      "postDate": "05/31/2018 15:21:53",
      "content": "<p>That's pretty awesome, good job !</p>",
      "rawMarkdown": "That's pretty awesome, good job !",
      "votes": null
    },
    {
      "id": "336608",
      "postDate": "06/01/2018 02:45:21",
      "content": "<p>How did you come to using 75 degrees as your angle of choice? </p>",
      "rawMarkdown": "How did you come to using 75 degrees as your angle of choice?",
      "votes": null
    },
    {
      "id": "337315",
      "postDate": "06/02/2018 14:29:34",
      "content": "<p>flatten: angle vs radius plot</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/337315/9561/angle_radius.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "flatten: angle vs radius plot\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/337315/9561/angle_radius.png",
      "votes": null
    },
    {
      "id": "337369",
      "postDate": "06/02/2018 17:14:14",
      "content": "<p>I think it is just an example. The method has to be done for each degree</p>",
      "rawMarkdown": "I think it is just an example. The method has to be done for each degree",
      "votes": null
    },
    {
      "id": "337483",
      "postDate": "06/03/2018 00:39:13",
      "content": "<p>Thank you for sharing your progress as you're climbing the leader board, best of luck! </p>",
      "rawMarkdown": "Thank you for sharing your progress as you're climbing the leader board, best of luck!",
      "votes": null
    },
    {
      "id": "337515",
      "postDate": "06/03/2018 04:12:25",
      "content": "<p>What do you mean by backfitting?</p>",
      "rawMarkdown": "What do you mean by backfitting?",
      "votes": null
    },
    {
      "id": "337588",
      "postDate": "06/03/2018 09:01:04",
      "content": "<p>@John Sweeney</p>\n\n<p>It is to determine the track parameters and extend the track.</p>\n\n<p>currently, my solution is:</p>\n\n<ol>\n<li><p>ensemble of \"different slicing / unrolling methods + DSCAN\"  : LB= 0.51</p></li>\n<li><p>line fitting and extension (backfitting) : LB= 0.55</p></li>\n</ol>\n\n<p>time to move beyond DBSCAN! I am applying 3d CNN for track reconstructions. The results are very promising. But there is a problem of making it scalable and efficient for such a large number of hits.</p>\n\n<p>Note that the track do not follow the helix equation exactly, hence there is is a limit to DSCAN methods. (e.g. it is not possible to unroll all  tracks to straight lines)</p>",
      "rawMarkdown": "John Sweeney\n\nIt is to determine the track parameters and extend the track.\n\ncurrently, my solution is:\n\n1. ensemble of \"different slicing / unrolling methods + DSCAN\"  : LB= 0.51\n\n2.  line fitting and extension (backfitting) : LB= 0.55\n\ntime to move beyond DBSCAN! I am applying 3d CNN for track reconstructions. The results are very promising. But there is a problem of making it scalable and efficient for such a large number of hits.\n\nNote that the track do not follow the helix equation exactly, hence there is is a limit to DSCAN methods. (e.g. it is not possible to unroll all  tracks to straight lines)",
      "votes": null
    },
    {
      "id": "337677",
      "postDate": "06/03/2018 13:28:48",
      "content": "<p>looking forward for the CNN update.  </p>",
      "rawMarkdown": "looking forward for the CNN update.",
      "votes": null
    },
    {
      "id": "337770",
      "postDate": "06/03/2018 17:54:55",
      "content": "<p>@Heng do you think PointCNN or point cloud methods in general are viable solutions ?</p>",
      "rawMarkdown": "Heng do you think PointCNN or point cloud methods in general are viable solutions ?",
      "votes": null
    },
    {
      "id": "337943",
      "postDate": "06/04/2018 05:43:09",
      "content": "<p>@AncientMonk</p>\n\n<p>initial results on my local validation set show good results. At least they are be used in part of the solution.</p>\n\n<p>\" PointCNN or point cloud methods \" ... maybe not necessary PointCNN , but a deep learning framework to deal with unstructured data, rather than regular grid data like image is viable i think.</p>",
      "rawMarkdown": "AncientMonk\n \ninitial results on my local validation set show good results. At least they are be used in part of the solution.\n\n\" PointCNN or point cloud methods \" ... maybe not necessary PointCNN , but a deep learning framework to deal with unstructured data, rather than regular grid data like image is viable i think.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 336376,
      "author_name": "gaarv1911",
      "author_url": "",
      "post_date": "05/31/2018 15:21:53",
      "content": "<p>That's pretty awesome, good job !</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 336608,
      "author_name": "anirudhk",
      "author_url": "",
      "post_date": "06/01/2018 02:45:21",
      "content": "<p>How did you come to using 75 degrees as your angle of choice? </p>",
      "votes": null,
      "replies": [
        {
          "id": 337369,
          "author_name": "darkdoval",
          "author_url": "",
          "post_date": "06/02/2018 17:14:14",
          "content": "<p>I think it is just an example. The method has to be done for each degree</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 337315,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/02/2018 14:29:34",
      "content": "<p>flatten: angle vs radius plot</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/337315/9561/angle_radius.png\" alt=\"enter image description here\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 337483,
          "author_name": "osa111",
          "author_url": "",
          "post_date": "06/03/2018 00:39:13",
          "content": "<p>Thank you for sharing your progress as you're climbing the leader board, best of luck! </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 337515,
      "author_name": "johnhsweeney",
      "author_url": "",
      "post_date": "06/03/2018 04:12:25",
      "content": "<p>What do you mean by backfitting?</p>",
      "votes": null,
      "replies": [
        {
          "id": 337588,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "06/03/2018 09:01:04",
          "content": "<p>@John Sweeney</p>\n\n<p>It is to determine the track parameters and extend the track.</p>\n\n<p>currently, my solution is:</p>\n\n<ol>\n<li><p>ensemble of \"different slicing / unrolling methods + DSCAN\"  : LB= 0.51</p></li>\n<li><p>line fitting and extension (backfitting) : LB= 0.55</p></li>\n</ol>\n\n<p>time to move beyond DBSCAN! I am applying 3d CNN for track reconstructions. The results are very promising. But there is a problem of making it scalable and efficient for such a large number of hits.</p>\n\n<p>Note that the track do not follow the helix equation exactly, hence there is is a limit to DSCAN methods. (e.g. it is not possible to unroll all  tracks to straight lines)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 337677,
          "author_name": "hireme",
          "author_url": "",
          "post_date": "06/03/2018 13:28:48",
          "content": "<p>looking forward for the CNN update.  </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 337770,
          "author_name": "warlord",
          "author_url": "",
          "post_date": "06/03/2018 17:54:55",
          "content": "<p>@Heng do you think PointCNN or point cloud methods in general are viable solutions ?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 337943,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "06/04/2018 05:43:09",
          "content": "<p>@AncientMonk</p>\n\n<p>initial results on my local validation set show good results. At least they are be used in part of the solution.</p>\n\n<p>\" PointCNN or point cloud methods \" ... maybe not necessary PointCNN , but a deep learning framework to deal with unstructured data, rather than regular grid data like image is viable i think.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "336192": "as shown below\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/336192/9547/Slide1.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/336192/9549/Slide3.png",
    "336376": "That's pretty awesome, good job !",
    "336608": "How did you come to using 75 degrees as your angle of choice?",
    "337315": "flatten: angle vs radius plot\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/337315/9561/angle_radius.png",
    "337369": "I think it is just an example. The method has to be done for each degree",
    "337483": "Thank you for sharing your progress as you're climbing the leader board, best of luck!",
    "337515": "What do you mean by backfitting?",
    "337588": "John Sweeney\n\nIt is to determine the track parameters and extend the track.\n\ncurrently, my solution is:\n\n1. ensemble of \"different slicing / unrolling methods + DSCAN\"  : LB= 0.51\n\n2.  line fitting and extension (backfitting) : LB= 0.55\n\ntime to move beyond DBSCAN! I am applying 3d CNN for track reconstructions. The results are very promising. But there is a problem of making it scalable and efficient for such a large number of hits.\n\nNote that the track do not follow the helix equation exactly, hence there is is a limit to DSCAN methods. (e.g. it is not possible to unroll all  tracks to straight lines)",
    "337677": "looking forward for the CNN update.",
    "337770": "Heng do you think PointCNN or point cloud methods in general are viable solutions ?",
    "337943": "AncientMonk\n \ninitial results on my local validation set show good results. At least they are be used in part of the solution.\n\n\" PointCNN or point cloud methods \" ... maybe not necessary PointCNN , but a deep learning framework to deal with unstructured data, rather than regular grid data like image is viable i think."
  },
  "source": "meta"
}