{
  "id": 60127,
  "title": "CNN unet segmentation results",
  "url": "/competitions/trackml-particle-identification/discussion/60127",
  "author_name": "",
  "post_date": "2018-06-30T13:37:48.706601500Z",
  "votes": 7,
  "comment_count": 4,
  "views": 0,
  "content": "<p>As attached:</p>\n\n<p>The projection to 2d image is to capture the \"relative location information (i.e. graph, topology)\". Each image pixel contains  (x,y,z) or (a,z/r) instaed of (r,gb).</p>\n\n<p>I borrow idea from \"splatnet\". “SPLATNet allows easy mapping of 2D information into 3D and vice-versa, resulting in a novel network architecture for joint processing of 3D point clouds and multi-view images.”</p>\n\n<p>How to use:</p>\n\n<ul>\n<li><p>project the 3d volume onto 2d image planes. The objective is to make the track as compact as possible (so that the whole track can be covered by CNN conv filter, i.e. within the receptive field).</p></li>\n<li><p>one input volume many have many projections, i.e. multi-view representation</p></li>\n<li><p>two separate 3d hit should NOT be projected completely onto  same 2d image location, else you will lose information</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9736/Slide6.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9737/Slide7.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9738/Slide8.png\" alt=\"enter image description here\"></p></li>\n</ul>\n\n<p><a href=\"https://news.developer.nvidia.com/nvidia-splatnet-research-paper-wins-a-major-cvpr-2018-award/\">https://news.developer.nvidia.com/nvidia-splatnet-research-paper-wins-a-major-cvpr-2018-award/</a></p>",
  "messages": [
    {
      "id": "350803",
      "postDate": "06/30/2018 13:37:48",
      "content": "<p>As attached:</p>\n\n<p>The projection to 2d image is to capture the \"relative location information (i.e. graph, topology)\". Each image pixel contains  (x,y,z) or (a,z/r) instaed of (r,gb).</p>\n\n<p>I borrow idea from \"splatnet\". “SPLATNet allows easy mapping of 2D information into 3D and vice-versa, resulting in a novel network architecture for joint processing of 3D point clouds and multi-view images.”</p>\n\n<p>How to use:</p>\n\n<ul>\n<li><p>project the 3d volume onto 2d image planes. The objective is to make the track as compact as possible (so that the whole track can be covered by CNN conv filter, i.e. within the receptive field).</p></li>\n<li><p>one input volume many have many projections, i.e. multi-view representation</p></li>\n<li><p>two separate 3d hit should NOT be projected completely onto  same 2d image location, else you will lose information</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9736/Slide6.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9737/Slide7.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9738/Slide8.png\" alt=\"enter image description here\"></p></li>\n</ul>\n\n<p><a href=\"https://news.developer.nvidia.com/nvidia-splatnet-research-paper-wins-a-major-cvpr-2018-award/\">https://news.developer.nvidia.com/nvidia-splatnet-research-paper-wins-a-major-cvpr-2018-award/</a></p>",
      "rawMarkdown": "As attached:\n\nThe projection to 2d image is to capture the \"relative location information (i.e. graph, topology)\". Each image pixel contains  (x,y,z) or (a,z/r) instaed of (r,gb).\n\nI borrow idea from \"splatnet\". “SPLATNet allows easy mapping of 2D information into 3D and vice-versa, resulting in a novel network architecture for joint processing of 3D point clouds and multi-view images.”\n\nHow to use:\n\n- project the 3d volume onto 2d image planes. The objective is to make the track as compact as possible (so that the whole track can be covered by CNN conv filter, i.e. within the receptive field).\n\n- one input volume many have many projections, i.e. multi-view representation\n\n- two separate 3d hit should NOT be projected completely onto  same 2d image location, else you will lose information\n\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n\n  ![enter image description here][3]\n\nhttps://news.developer.nvidia.com/nvidia-splatnet-research-paper-wins-a-major-cvpr-2018-award/\n\n \n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9736/Slide6.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9737/Slide7.png\n  [3]: https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9738/Slide8.png",
      "votes": null
    },
    {
      "id": "350839",
      "postDate": "06/30/2018 15:12:26",
      "content": "<p>Awesome article. </p>",
      "rawMarkdown": "Awesome article.",
      "votes": null
    },
    {
      "id": "350885",
      "postDate": "06/30/2018 17:02:23",
      "content": "<p>example code and train results</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350885/9739/20600.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350885/9743/27800.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "example code and train results\n\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/350885/9739/20600.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/350885/9743/27800.png",
      "votes": null
    },
    {
      "id": "350889",
      "postDate": "06/30/2018 17:19:15",
      "content": "<p>for future better efficiency, you may want to google and check papers on:</p>\n\n<ul>\n<li><p>deformable conv : <a href=\"https://arxiv.org/abs/1703.06211\">https://arxiv.org/abs/1703.06211</a></p></li>\n<li><p>sparse conv: <a href=\"https://github.com/facebookresearch/SparseConvNet\">https://github.com/facebookresearch/SparseConvNet</a></p></li>\n<li><p>3d conv: <a href=\"https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py\">https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py</a></p></li>\n</ul>",
      "rawMarkdown": "for future better efficiency, you may want to google and check papers on:\n\n -  deformable conv : https://arxiv.org/abs/1703.06211\n\n - sparse conv: https://github.com/facebookresearch/SparseConvNet\n\n - 3d conv: https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py",
      "votes": null
    },
    {
      "id": "350969",
      "postDate": "06/30/2018 23:08:48",
      "content": "<p>Very cool! Thanks for sharing!</p>",
      "rawMarkdown": "Very cool! Thanks for sharing!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 350839,
      "author_name": "pavansanagapati",
      "author_url": "",
      "post_date": "06/30/2018 15:12:26",
      "content": "<p>Awesome article. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 350885,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/30/2018 17:02:23",
      "content": "<p>example code and train results</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350885/9739/20600.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350885/9743/27800.png\" alt=\"enter image description here\"></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 350889,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/30/2018 17:19:15",
      "content": "<p>for future better efficiency, you may want to google and check papers on:</p>\n\n<ul>\n<li><p>deformable conv : <a href=\"https://arxiv.org/abs/1703.06211\">https://arxiv.org/abs/1703.06211</a></p></li>\n<li><p>sparse conv: <a href=\"https://github.com/facebookresearch/SparseConvNet\">https://github.com/facebookresearch/SparseConvNet</a></p></li>\n<li><p>3d conv: <a href=\"https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py\">https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py</a></p></li>\n</ul>",
      "votes": null,
      "replies": []
    },
    {
      "id": 350969,
      "author_name": "blondinka",
      "author_url": "",
      "post_date": "06/30/2018 23:08:48",
      "content": "<p>Very cool! Thanks for sharing!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "350803": "As attached:\n\nThe projection to 2d image is to capture the \"relative location information (i.e. graph, topology)\". Each image pixel contains  (x,y,z) or (a,z/r) instaed of (r,gb).\n\nI borrow idea from \"splatnet\". “SPLATNet allows easy mapping of 2D information into 3D and vice-versa, resulting in a novel network architecture for joint processing of 3D point clouds and multi-view images.”\n\nHow to use:\n\n- project the 3d volume onto 2d image planes. The objective is to make the track as compact as possible (so that the whole track can be covered by CNN conv filter, i.e. within the receptive field).\n\n- one input volume many have many projections, i.e. multi-view representation\n\n- two separate 3d hit should NOT be projected completely onto  same 2d image location, else you will lose information\n\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n\n  ![enter image description here][3]\n\nhttps://news.developer.nvidia.com/nvidia-splatnet-research-paper-wins-a-major-cvpr-2018-award/\n\n \n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9736/Slide6.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9737/Slide7.png\n  [3]: https://storage.googleapis.com/kaggle-forum-message-attachments/350803/9738/Slide8.png",
    "350839": "Awesome article.",
    "350885": "example code and train results\n\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/350885/9739/20600.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/350885/9743/27800.png",
    "350889": "for future better efficiency, you may want to google and check papers on:\n\n -  deformable conv : https://arxiv.org/abs/1703.06211\n\n - sparse conv: https://github.com/facebookresearch/SparseConvNet\n\n - 3d conv: https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py",
    "350969": "Very cool! Thanks for sharing!"
  },
  "source": "meta"
}