{
  "id": 391892,
  "title": "31st place solution",
  "url": "/competitions/nfl-player-contact-detection/discussion/391892",
  "author_name": "tomofumin",
  "post_date": "2023-03-02T22:24:27.136000",
  "votes": 13,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Thanks to organizers and other participant for this exciting competition!<br>\nI really learned a lot from this competition and enjoyed it.<br>\nMy leaderboard rank is low, but I write down my solution as an example of not good one.</p>\n<h2>Overview of my solution</h2>\n<p>My pipeline consists of:</p>\n<ul>\n<li>2.5D CNN classification(7 class)</li>\n<li>xgboost (tracking data and 2.5D CNN output)</li>\n</ul>\n<h2>2.5D CNN classification</h2>\n<ul>\n<li>The output of the model was classified into the following 7 class:</li>\n</ul>\n<pre><code>no contact\nplayer contact\nground\nprevious player contact\npost player contact\nprevious ground\npost ground\n</code></pre>\n<p>I decided \"previous\" and \"post\" are two steps before and after the contact occurs.</p>\n<ul>\n<li>Around 3 frames were used in the model, and predicted with a total of 6 images by Sideline and Endzone.</li>\n<li>The size of the helmet changed the clip size from the image.((256,256),(128,128))</li>\n<li>I wrote the boundy box of the helmet directly on the image.</li>\n<li>Besides the images, only G_flug and distance were included in the model features.</li>\n<li>I choose efficientnet b2 pretrained,leaning rate 1e-3, Adam for learning.</li>\n<li>matthews corrcoef score achieved 0.59.</li>\n</ul>\n<h2>xgboost</h2>\n<ul>\n<li>The following features were added to the baseline:</li>\n<li>The tracking data around 10steps.</li>\n<li>The predictions value of the 2.5D CNN model around 20steps.</li>\n<li>whether the image could be retrieved by dataloader.</li>\n<li>The difference in distance between the helmets of player1,2, which is divided by the average of the helmet width and height.</li>\n<li>CV 0.740 / LB 0.746 / Private LB 0.742</li>\n</ul>",
  "messages": [
    {
      "id": 2166644,
      "postDate": "2023-03-02T22:24:27.137Z",
      "content": "<p>Thanks to organizers and other participant for this exciting competition!<br>\nI really learned a lot from this competition and enjoyed it.<br>\nMy leaderboard rank is low, but I write down my solution as an example of not good one.</p>\n<h2>Overview of my solution</h2>\n<p>My pipeline consists of:</p>\n<ul>\n<li>2.5D CNN classification(7 class)</li>\n<li>xgboost (tracking data and 2.5D CNN output)</li>\n</ul>\n<h2>2.5D CNN classification</h2>\n<ul>\n<li>The output of the model was classified into the following 7 class:</li>\n</ul>\n<pre><code>no contact\nplayer contact\nground\nprevious player contact\npost player contact\nprevious ground\npost ground\n</code></pre>\n<p>I decided \"previous\" and \"post\" are two steps before and after the contact occurs.</p>\n<ul>\n<li>Around 3 frames were used in the model, and predicted with a total of 6 images by Sideline and Endzone.</li>\n<li>The size of the helmet changed the clip size from the image.((256,256),(128,128))</li>\n<li>I wrote the boundy box of the helmet directly on the image.</li>\n<li>Besides the images, only G_flug and distance were included in the model features.</li>\n<li>I choose efficientnet b2 pretrained,leaning rate 1e-3, Adam for learning.</li>\n<li>matthews corrcoef score achieved 0.59.</li>\n</ul>\n<h2>xgboost</h2>\n<ul>\n<li>The following features were added to the baseline:</li>\n<li>The tracking data around 10steps.</li>\n<li>The predictions value of the 2.5D CNN model around 20steps.</li>\n<li>whether the image could be retrieved by dataloader.</li>\n<li>The difference in distance between the helmets of player1,2, which is divided by the average of the helmet width and height.</li>\n<li>CV 0.740 / LB 0.746 / Private LB 0.742</li>\n</ul>",
      "rawMarkdown": "Thanks to organizers and other participant for this exciting competition!\nI really learned a lot from this competition and enjoyed it.\nMy leaderboard rank is low, but I write down my solution as an example of not good one.\n## Overview of my solution\nMy pipeline consists of:\n- 2.5D CNN classification(7 class)\n- xgboost (tracking data and 2.5D CNN output)\n## 2.5D CNN classification\n- The output of the model was classified into the following 7 class:\n```\nno contact\nplayer contact\nground\nprevious player contact\npost player contact\nprevious ground\npost ground\n```\nI decided \"previous\" and \"post\" are two steps before and after the contact occurs.\n- Around 3 frames were used in the model, and predicted with a total of 6 images by Sideline and Endzone.\n- The size of the helmet changed the clip size from the image.((256,256),(128,128))\n- I wrote the boundy box of the helmet directly on the image.\n- Besides the images, only G_flug and distance were included in the model features.\n- I choose efficientnet b2 pretrained,leaning rate 1e-3, Adam for learning.\n- matthews corrcoef score achieved 0.59.\n\n##  xgboost\n- The following features were added to the baseline:\n- The tracking data around 10steps.\n- The predictions value of the 2.5D CNN model around 20steps.\n- whether the image could be retrieved by dataloader.\n- The difference in distance between the helmets of player1,2, which is divided by the average of the helmet width and height.\n- CV 0.740 / LB 0.746 / Private LB 0.742",
      "votes": 13
    },
    {
      "id": 2166688,
      "postDate": "2023-03-02T23:53:04.123Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 2166688,
      "author_name": "",
      "author_url": "",
      "post_date": "2023-03-02T23:53:04.123000",
      "content": "",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2166644": "Thanks to organizers and other participant for this exciting competition!\nI really learned a lot from this competition and enjoyed it.\nMy leaderboard rank is low, but I write down my solution as an example of not good one.\n## Overview of my solution\nMy pipeline consists of:\n- 2.5D CNN classification(7 class)\n- xgboost (tracking data and 2.5D CNN output)\n## 2.5D CNN classification\n- The output of the model was classified into the following 7 class:\n```\nno contact\nplayer contact\nground\nprevious player contact\npost player contact\nprevious ground\npost ground\n```\nI decided \"previous\" and \"post\" are two steps before and after the contact occurs.\n- Around 3 frames were used in the model, and predicted with a total of 6 images by Sideline and Endzone.\n- The size of the helmet changed the clip size from the image.((256,256),(128,128))\n- I wrote the boundy box of the helmet directly on the image.\n- Besides the images, only G_flug and distance were included in the model features.\n- I choose efficientnet b2 pretrained,leaning rate 1e-3, Adam for learning.\n- matthews corrcoef score achieved 0.59.\n\n##  xgboost\n- The following features were added to the baseline:\n- The tracking data around 10steps.\n- The predictions value of the 2.5D CNN model around 20steps.\n- whether the image could be retrieved by dataloader.\n- The difference in distance between the helmets of player1,2, which is divided by the average of the helmet width and height.\n- CV 0.740 / LB 0.746 / Private LB 0.742",
    "2166688": ""
  }
}