{
  "id": 239884,
  "title": "Super simple and super effective trick - 13th place",
  "url": "/competitions/indoor-location-navigation/discussion/239884",
  "author_name": "",
  "post_date": "2021-05-18T00:34:30.414868Z",
  "votes": 21,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Thank you all for the competition!</p>\n<p>Organizers provided us with functions which enabled computing vectors of movement (x and y coordinates). Average error ((x_hat - x) ^ 2 + (y_hat - y) ^ 2) ^ 0.5 was <strong>2.67m</strong>. I have build a MLP which corrected this calculation and brought the error down to <strong>1.86m</strong>. The only input was: <strong>building</strong> one hot vector, <strong>floor</strong> one hot vector and <strong>raw calculations</strong>! Even without building and floor (with just row calculations!) the error could be brought down by around 25%.</p>\n<p>It boosted the cost minimization. Later I have used tetris-like fitting paths into hallways. I am going to share more in the following days.</p>",
  "messages": [
    {
      "id": "1312273",
      "postDate": "05/18/2021 00:34:30",
      "content": "<p>Thank you all for the competition!</p>\n<p>Organizers provided us with functions which enabled computing vectors of movement (x and y coordinates). Average error ((x_hat - x) ^ 2 + (y_hat - y) ^ 2) ^ 0.5 was <strong>2.67m</strong>. I have build a MLP which corrected this calculation and brought the error down to <strong>1.86m</strong>. The only input was: <strong>building</strong> one hot vector, <strong>floor</strong> one hot vector and <strong>raw calculations</strong>! Even without building and floor (with just row calculations!) the error could be brought down by around 25%.</p>\n<p>It boosted the cost minimization. Later I have used tetris-like fitting paths into hallways. I am going to share more in the following days.</p>",
      "rawMarkdown": "Thank you all for the competition!\n\nOrganizers provided us with functions which enabled computing vectors of movement (x and y coordinates). Average error ((x_hat - x) ^ 2 + (y_hat - y) ^ 2) ^ 0.5 was **2.67m**. I have build a MLP which corrected this calculation and brought the error down to **1.86m**. The only input was: **building** one hot vector, **floor** one hot vector and **raw calculations**! Even without building and floor (with just row calculations!) the error could be brought down by around 25%.\n\nIt boosted the cost minimization. Later I have used tetris-like fitting paths into hallways. I am going to share more in the following days.",
      "votes": null
    },
    {
      "id": "1312275",
      "postDate": "05/18/2021 00:38:11",
      "content": "<p>yeah, I'm using the same trick. I made it to <code>1.58 m</code> in the end. In case anyone is wondering, Olaf is talking about <code>delta x &amp; y</code> in this model.</p>",
      "rawMarkdown": "yeah, I'm using the same trick. I made it to `1.58 m` in the end. In case anyone is wondering, Olaf is talking about `delta x & y` in this model.",
      "votes": null
    },
    {
      "id": "1312277",
      "postDate": "05/18/2021 00:39:26",
      "content": "<p>I made it till 1.7m, but by combining the MLP with RNN and IMU data. I couldn't successfully predict delta using only IMU data, raw calculations helped a lot.</p>",
      "rawMarkdown": "I made it till 1.7m, but by combining the MLP with RNN and IMU data. I couldn't successfully predict delta using only IMU data, raw calculations helped a lot.",
      "votes": null
    },
    {
      "id": "1312278",
      "postDate": "05/18/2021 00:42:52",
      "content": "<p>Great Trick! Our team didn't notice it. </p>",
      "rawMarkdown": "Great Trick! Our team didn't notice it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1312275,
      "author_name": "jiweiliu",
      "author_url": "",
      "post_date": "05/18/2021 00:38:11",
      "content": "<p>yeah, I'm using the same trick. I made it to <code>1.58 m</code> in the end. In case anyone is wondering, Olaf is talking about <code>delta x &amp; y</code> in this model.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1312277,
          "author_name": "olaf2000",
          "author_url": "",
          "post_date": "05/18/2021 00:39:26",
          "content": "<p>I made it till 1.7m, but by combining the MLP with RNN and IMU data. I couldn't successfully predict delta using only IMU data, raw calculations helped a lot.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1312278,
      "author_name": "mamasinkgs",
      "author_url": "",
      "post_date": "05/18/2021 00:42:52",
      "content": "<p>Great Trick! Our team didn't notice it. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1312273": "Thank you all for the competition!\n\nOrganizers provided us with functions which enabled computing vectors of movement (x and y coordinates). Average error ((x_hat - x) ^ 2 + (y_hat - y) ^ 2) ^ 0.5 was **2.67m**. I have build a MLP which corrected this calculation and brought the error down to **1.86m**. The only input was: **building** one hot vector, **floor** one hot vector and **raw calculations**! Even without building and floor (with just row calculations!) the error could be brought down by around 25%.\n\nIt boosted the cost minimization. Later I have used tetris-like fitting paths into hallways. I am going to share more in the following days.",
    "1312275": "yeah, I'm using the same trick. I made it to `1.58 m` in the end. In case anyone is wondering, Olaf is talking about `delta x & y` in this model.",
    "1312277": "I made it till 1.7m, but by combining the MLP with RNN and IMU data. I couldn't successfully predict delta using only IMU data, raw calculations helped a lot.",
    "1312278": "Great Trick! Our team didn't notice it."
  },
  "source": "meta"
}