{
  "id": 58244,
  "title": "Track seed filtering and deep learning",
  "url": "/competitions/trackml-particle-identification/discussion/58244",
  "author_name": "",
  "post_date": "2018-06-05T00:51:47.543245700Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi,</p>\n\n<p>I was looking into several ways of how to find seed tracks as a starting point to allow more complex networks (LSTM, CNN, graph network, etc) to be used on the data. My current approach is a MLP that predicts the probability that two hits from adjacent layer_ids in volume 8 (The volume closest to the center of the detector where the tracks originate) are part of the same track. Aside from this being an inefficent method of track seed filtering, the results from this approach are pretty bad. </p>\n\n<p>Are there any other approaches to find track seeds that would be more efficient and give better results?</p>\n\n<p>Thanks</p>",
  "messages": [
    {
      "id": "338402",
      "postDate": "06/05/2018 00:51:47",
      "content": "<p>Hi,</p>\n\n<p>I was looking into several ways of how to find seed tracks as a starting point to allow more complex networks (LSTM, CNN, graph network, etc) to be used on the data. My current approach is a MLP that predicts the probability that two hits from adjacent layer_ids in volume 8 (The volume closest to the center of the detector where the tracks originate) are part of the same track. Aside from this being an inefficent method of track seed filtering, the results from this approach are pretty bad. </p>\n\n<p>Are there any other approaches to find track seeds that would be more efficient and give better results?</p>\n\n<p>Thanks</p>",
      "rawMarkdown": "Hi,\n\nI was looking into several ways of how to find seed tracks as a starting point to allow more complex networks (LSTM, CNN, graph network, etc) to be used on the data. My current approach is a MLP that predicts the probability that two hits from adjacent layer_ids in volume 8 (The volume closest to the center of the detector where the tracks originate) are part of the same track. Aside from this being an inefficent method of track seed filtering, the results from this approach are pretty bad. \n\nAre there any other approaches to find track seeds that would be more efficient and give better results?\n\nThanks",
      "votes": null
    },
    {
      "id": "338473",
      "postDate": "06/05/2018 06:20:28",
      "content": "<p>I've been assuming that we could use dbscan as discussed extensively on the board to source seed tracks...</p>",
      "rawMarkdown": "I've been assuming that we could use dbscan as discussed extensively on the board to source seed tracks...",
      "votes": null
    },
    {
      "id": "339611",
      "postDate": "06/07/2018 07:48:06",
      "content": "<p><a href=\"http://slideplayer.com/slide/8811271/\">http://slideplayer.com/slide/8811271/</a>  </p>\n\n<p><a href=\"http://folk.uio.no/ares/FYS4550/rudiVCI.pdf\">http://folk.uio.no/ares/FYS4550/rudiVCI.pdf</a></p>\n\n<p><img src=\"http://slideplayer.com/8811271/26/images/14/Tracklet+Search+Projection+Combinatorial+algorithm+too+expensive.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"http://slideplayer.com/8811271/26/images/13/Tracklet+Search%3A+Principles.jpg\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "http://slideplayer.com/slide/8811271/  \n\nhttp://folk.uio.no/ares/FYS4550/rudiVCI.pdf\n\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n\n\n  [1]: http://slideplayer.com/8811271/26/images/14/Tracklet+Search+Projection+Combinatorial+algorithm+too+expensive.jpg\n  [2]: http://slideplayer.com/8811271/26/images/13/Tracklet+Search%3A+Principles.jpg",
      "votes": null
    },
    {
      "id": "339818",
      "postDate": "06/07/2018 18:09:35",
      "content": "<p>Hi,</p>\n\n<p>Would the triplet generation sample code <a href=\"https://www.kaggle.com/c/trackml-particle-identification/discussion/58292#339779\">here</a> be useful in narrowing down all possible combinations of hits to find track seeds. Also, could it be useful to find hits that are close to the next hit on a reference track seed as negative examples to train a MLP to differentiate between actual and fake track seeds?</p>\n\n<p>Thanks </p>",
      "rawMarkdown": "Hi,\n\nWould the triplet generation sample code [here][1] be useful in narrowing down all possible combinations of hits to find track seeds. Also, could it be useful to find hits that are close to the next hit on a reference track seed as negative examples to train a MLP to differentiate between actual and fake track seeds?\n\nThanks \n\n[1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/58292#339779",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 338473,
      "author_name": "johnhsweeney",
      "author_url": "",
      "post_date": "06/05/2018 06:20:28",
      "content": "<p>I've been assuming that we could use dbscan as discussed extensively on the board to source seed tracks...</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 339611,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/07/2018 07:48:06",
      "content": "<p><a href=\"http://slideplayer.com/slide/8811271/\">http://slideplayer.com/slide/8811271/</a>  </p>\n\n<p><a href=\"http://folk.uio.no/ares/FYS4550/rudiVCI.pdf\">http://folk.uio.no/ares/FYS4550/rudiVCI.pdf</a></p>\n\n<p><img src=\"http://slideplayer.com/8811271/26/images/14/Tracklet+Search+Projection+Combinatorial+algorithm+too+expensive.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"http://slideplayer.com/8811271/26/images/13/Tracklet+Search%3A+Principles.jpg\" alt=\"enter image description here\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 339818,
          "author_name": "bkkaggle",
          "author_url": "",
          "post_date": "06/07/2018 18:09:35",
          "content": "<p>Hi,</p>\n\n<p>Would the triplet generation sample code <a href=\"https://www.kaggle.com/c/trackml-particle-identification/discussion/58292#339779\">here</a> be useful in narrowing down all possible combinations of hits to find track seeds. Also, could it be useful to find hits that are close to the next hit on a reference track seed as negative examples to train a MLP to differentiate between actual and fake track seeds?</p>\n\n<p>Thanks </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "338402": "Hi,\n\nI was looking into several ways of how to find seed tracks as a starting point to allow more complex networks (LSTM, CNN, graph network, etc) to be used on the data. My current approach is a MLP that predicts the probability that two hits from adjacent layer_ids in volume 8 (The volume closest to the center of the detector where the tracks originate) are part of the same track. Aside from this being an inefficent method of track seed filtering, the results from this approach are pretty bad. \n\nAre there any other approaches to find track seeds that would be more efficient and give better results?\n\nThanks",
    "338473": "I've been assuming that we could use dbscan as discussed extensively on the board to source seed tracks...",
    "339611": "http://slideplayer.com/slide/8811271/  \n\nhttp://folk.uio.no/ares/FYS4550/rudiVCI.pdf\n\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n\n\n  [1]: http://slideplayer.com/8811271/26/images/14/Tracklet+Search+Projection+Combinatorial+algorithm+too+expensive.jpg\n  [2]: http://slideplayer.com/8811271/26/images/13/Tracklet+Search%3A+Principles.jpg",
    "339818": "Hi,\n\nWould the triplet generation sample code [here][1] be useful in narrowing down all possible combinations of hits to find track seeds. Also, could it be useful to find hits that are close to the next hit on a reference track seed as negative examples to train a MLP to differentiate between actual and fake track seeds?\n\nThanks \n\n[1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/58292#339779"
  },
  "source": "meta"
}