{
  "id": 300129,
  "title": "how to union  the detected bboxes  based on IoU?",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/300129",
  "author_name": "",
  "post_date": "2022-01-11T09:10:39.701575200Z",
  "votes": 1,
  "comment_count": 10,
  "views": 0,
  "content": "<p>not the way like  weighed_box_fusion,</p>\n<p>but like union of disjoint set?   if two boxes IoU over threshold, we take the bigger confidence one,</p>\n<p>otherwise, we keep the box.  </p>\n<p>anyone try that?</p>",
  "messages": [
    {
      "id": "1645821",
      "postDate": "01/11/2022 09:10:39",
      "content": "<p>not the way like  weighed_box_fusion,</p>\n<p>but like union of disjoint set?   if two boxes IoU over threshold, we take the bigger confidence one,</p>\n<p>otherwise, we keep the box.  </p>\n<p>anyone try that?</p>",
      "rawMarkdown": "not the way like  weighed_box_fusion,\n\nbut like union of disjoint set?   if two boxes IoU over threshold, we take the bigger confidence one,\n\notherwise, we keep the box.  \n\nanyone try that?",
      "votes": null
    },
    {
      "id": "1645989",
      "postDate": "01/11/2022 11:58:11",
      "content": "<p>Isn't that non-max suppression (NMS)?</p>",
      "rawMarkdown": "Isn't that non-max suppression (NMS)?",
      "votes": null
    },
    {
      "id": "1646027",
      "postDate": "01/11/2022 12:37:12",
      "content": "<p>here is the photos using NMS, not what I expected.</p>\n<p><a href=\"https://drive.google.com/drive/folders/1B4IrAKNBAXLKAPSR_jooPpkjN09ebsGx?usp=sharing\" target=\"_blank\">https://drive.google.com/drive/folders/1B4IrAKNBAXLKAPSR_jooPpkjN09ebsGx?usp=sharing</a></p>",
      "rawMarkdown": "here is the photos using NMS, not what I expected.\n\nhttps://drive.google.com/drive/folders/1B4IrAKNBAXLKAPSR_jooPpkjN09ebsGx?usp=sharing",
      "votes": null
    },
    {
      "id": "1646044",
      "postDate": "01/11/2022 12:53:34",
      "content": "<p>You could try to use the WBF bounding boxes enabling approach. Take a look to <a href=\"https://www.kaggle.com/shonenkov/wbf-approach-for-ensemble\" target=\"_blank\">detailed dicussion</a></p>",
      "rawMarkdown": "You could try to use the WBF bounding boxes enabling approach. Take a look to [detailed dicussion](https://www.kaggle.com/shonenkov/wbf-approach-for-ensemble)",
      "votes": null
    },
    {
      "id": "1646256",
      "postDate": "01/11/2022 15:29:23",
      "content": "<p>I think also what you describe is NMS </p>",
      "rawMarkdown": "I think also what you describe is NMS",
      "votes": null
    },
    {
      "id": "1646338",
      "postDate": "01/11/2022 16:36:38",
      "content": "<p>thx for reply. I will take a look. </p>",
      "rawMarkdown": "thx for reply. I will take a look.",
      "votes": null
    },
    {
      "id": "1647108",
      "postDate": "01/12/2022 10:15:11",
      "content": "<p>The result is not correct obviously,<br>\ncheck your code or NMS usage</p>",
      "rawMarkdown": "The result is not correct obviously,\ncheck your code or NMS usage",
      "votes": null
    },
    {
      "id": "1648103",
      "postDate": "01/13/2022 06:21:08",
      "content": "<p>thx, I am looking into it.  previously just copied without much thinking.</p>",
      "rawMarkdown": "thx, I am looking into it.  previously just copied without much thinking.",
      "votes": null
    },
    {
      "id": "1648333",
      "postDate": "01/13/2022 11:06:22",
      "content": "<p>based on Hengck23's result,  WBF ensemble for single model no improvement on LB.</p>",
      "rawMarkdown": "based on Hengck23's result,  WBF ensemble for single model no improvement on LB.",
      "votes": null
    },
    {
      "id": "1648353",
      "postDate": "01/13/2022 11:25:24",
      "content": "<p>my WBF ensemble is based on fold0,1,2.<br>\nif there FPR is not too high, there is an improvement for the ensemble.</p>\n<p>however, if the FPR is high for single model, ensemble will result in higher FPR (more than TP), then your f2 score will drop</p>",
      "rawMarkdown": "my WBF ensemble is based on fold0,1,2.\nif there FPR is not too high, there is an improvement for the ensemble.\n\nhowever, if the FPR is high for single model, ensemble will result in higher FPR (more than TP), then your f2 score will drop",
      "votes": null
    },
    {
      "id": "1648797",
      "postDate": "01/13/2022 17:39:19",
      "content": "<p>thx for your reply.   I just give up experiments.</p>",
      "rawMarkdown": "thx for your reply.   I just give up experiments.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1645989,
      "author_name": "hannes82",
      "author_url": "",
      "post_date": "01/11/2022 11:58:11",
      "content": "<p>Isn't that non-max suppression (NMS)?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1646027,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "01/11/2022 12:37:12",
      "content": "<p>here is the photos using NMS, not what I expected.</p>\n<p><a href=\"https://drive.google.com/drive/folders/1B4IrAKNBAXLKAPSR_jooPpkjN09ebsGx?usp=sharing\" target=\"_blank\">https://drive.google.com/drive/folders/1B4IrAKNBAXLKAPSR_jooPpkjN09ebsGx?usp=sharing</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1647108,
          "author_name": "atom1231",
          "author_url": "",
          "post_date": "01/12/2022 10:15:11",
          "content": "<p>The result is not correct obviously,<br>\ncheck your code or NMS usage</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1648103,
          "author_name": "dragonzhang",
          "author_url": "",
          "post_date": "01/13/2022 06:21:08",
          "content": "<p>thx, I am looking into it.  previously just copied without much thinking.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1646044,
      "author_name": "meowmeowmeowmeowmeow",
      "author_url": "",
      "post_date": "01/11/2022 12:53:34",
      "content": "<p>You could try to use the WBF bounding boxes enabling approach. Take a look to <a href=\"https://www.kaggle.com/shonenkov/wbf-approach-for-ensemble\" target=\"_blank\">detailed dicussion</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1646338,
          "author_name": "dragonzhang",
          "author_url": "",
          "post_date": "01/11/2022 16:36:38",
          "content": "<p>thx for reply. I will take a look. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1646256,
      "author_name": "imeintanis",
      "author_url": "",
      "post_date": "01/11/2022 15:29:23",
      "content": "<p>I think also what you describe is NMS </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1648333,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "01/13/2022 11:06:22",
      "content": "<p>based on Hengck23's result,  WBF ensemble for single model no improvement on LB.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1648353,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "01/13/2022 11:25:24",
          "content": "<p>my WBF ensemble is based on fold0,1,2.<br>\nif there FPR is not too high, there is an improvement for the ensemble.</p>\n<p>however, if the FPR is high for single model, ensemble will result in higher FPR (more than TP), then your f2 score will drop</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1648797,
          "author_name": "dragonzhang",
          "author_url": "",
          "post_date": "01/13/2022 17:39:19",
          "content": "<p>thx for your reply.   I just give up experiments.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1645821": "not the way like  weighed_box_fusion,\n\nbut like union of disjoint set?   if two boxes IoU over threshold, we take the bigger confidence one,\n\notherwise, we keep the box.  \n\nanyone try that?",
    "1645989": "Isn't that non-max suppression (NMS)?",
    "1646027": "here is the photos using NMS, not what I expected.\n\nhttps://drive.google.com/drive/folders/1B4IrAKNBAXLKAPSR_jooPpkjN09ebsGx?usp=sharing",
    "1646044": "You could try to use the WBF bounding boxes enabling approach. Take a look to [detailed dicussion](https://www.kaggle.com/shonenkov/wbf-approach-for-ensemble)",
    "1646256": "I think also what you describe is NMS",
    "1646338": "thx for reply. I will take a look.",
    "1647108": "The result is not correct obviously,\ncheck your code or NMS usage",
    "1648103": "thx, I am looking into it.  previously just copied without much thinking.",
    "1648333": "based on Hengck23's result,  WBF ensemble for single model no improvement on LB.",
    "1648353": "my WBF ensemble is based on fold0,1,2.\nif there FPR is not too high, there is an improvement for the ensemble.\n\nhowever, if the FPR is high for single model, ensemble will result in higher FPR (more than TP), then your f2 score will drop",
    "1648797": "thx for your reply.   I just give up experiments."
  },
  "source": "meta"
}