{
  "id": 307640,
  "title": "Simple trick for AutoWBF",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/307640",
  "author_name": "Lukasz Borecki",
  "post_date": "2022-02-15T04:37:31.595000",
  "votes": 10,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I want to share one of the tricks we found to be working with wbf. After increasing IOU at some point we used 0.9 and 0.95 there were multiply overlaping boxes on SF. We overwhelmed NMS with purpose to switch it to WBF. After this manuevre score goes up by 0.03 for free (used in final submission) (In PVT LB it goes up by 0.007). </p>\n<p>Follow up was feature engineering to count number of overlaping boxes. Usually the greater number  we could assume its likely TP. We used that to remove some FP but gain was only 0.002, (didn't use it in final sub)</p>\n<p>------------ Visualized concept ------------------<br>\n<img src=\"https://i.ibb.co/pP0y6br/autowbf1.png\" alt=\"\"></p>\n<p><img src=\"https://i.ibb.co/rHHwBMB/autowbf2.png\" alt=\"\"></p>",
  "messages": [
    {
      "id": 1690734,
      "postDate": "2022-02-15T04:37:31.597Z",
      "content": "<p>I want to share one of the tricks we found to be working with wbf. After increasing IOU at some point we used 0.9 and 0.95 there were multiply overlaping boxes on SF. We overwhelmed NMS with purpose to switch it to WBF. After this manuevre score goes up by 0.03 for free (used in final submission) (In PVT LB it goes up by 0.007). </p>\n<p>Follow up was feature engineering to count number of overlaping boxes. Usually the greater number  we could assume its likely TP. We used that to remove some FP but gain was only 0.002, (didn't use it in final sub)</p>\n<p>------------ Visualized concept ------------------<br>\n<img src=\"https://i.ibb.co/pP0y6br/autowbf1.png\" alt=\"\"></p>\n<p><img src=\"https://i.ibb.co/rHHwBMB/autowbf2.png\" alt=\"\"></p>",
      "rawMarkdown": "I want to share one of the tricks we found to be working with wbf. After increasing IOU at some point we used 0.9 and 0.95 there were multiply overlaping boxes on SF. We overwhelmed NMS with purpose to switch it to WBF. After this manuevre score goes up by 0.03 for free (used in final submission) (In PVT LB it goes up by 0.007). \n\nFollow up was feature engineering to count number of overlaping boxes. Usually the greater number  we could assume its likely TP. We used that to remove some FP but gain was only 0.002, (didn't use it in final sub)\n\n------------ Visualized concept ------------------\n![](https://i.ibb.co/pP0y6br/autowbf1.png)\n\n![](https://i.ibb.co/rHHwBMB/autowbf2.png)",
      "votes": 10
    },
    {
      "id": 1692581,
      "postDate": "2022-02-16T06:06:26.403Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1692581,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-02-16T06:06:26.403000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1690734": "I want to share one of the tricks we found to be working with wbf. After increasing IOU at some point we used 0.9 and 0.95 there were multiply overlaping boxes on SF. We overwhelmed NMS with purpose to switch it to WBF. After this manuevre score goes up by 0.03 for free (used in final submission) (In PVT LB it goes up by 0.007). \n\nFollow up was feature engineering to count number of overlaping boxes. Usually the greater number  we could assume its likely TP. We used that to remove some FP but gain was only 0.002, (didn't use it in final sub)\n\n------------ Visualized concept ------------------\n![](https://i.ibb.co/pP0y6br/autowbf1.png)\n\n![](https://i.ibb.co/rHHwBMB/autowbf2.png)",
    "1692581": ""
  }
}