{
  "id": 247535,
  "title": "cufused with hyp when i use yolo",
  "url": "/competitions/siim-covid19-detection/discussion/247535",
  "author_name": "",
  "post_date": "2021-06-20T04:20:03.760055300Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>when i use yolo and run test.py ,</p>\n<pre><code>python test.py --img 512 --batch 32 --conf 0.2 --iou 0.1:mAP@0.5=0.918\npython test.py --img 512 --batch 32 --conf 0.15 --iou 0.1:mAP@0.5=0.824\npython test.py --img 512 --batch 32 --conf 0.1 --iou 0.1:mAP@0.5=0.750\npython test.py --img 512 --batch 32 --conf 0.01 --iou 0.1:mAP@0.5=0.589\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.1:mAP@0.5=0.526\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.3:mAP@0.5=0.499\npython test.py --img 512 --batch 32 ---conf 0.001 --iou 0.65:mAP@0.5=0.471\n</code></pre>\n<p>I don't know why mAP is so sensitive to these param, because i 'm a new to object detection.</p>\n<pre><code>names:\n- 0. opacity\nnc: 1\ntrain: ~/512img600study/fold0_tra.txt\nval: ~/512img600study/fold0_val.txt\n</code></pre>\n<p>i use these for --data</p>",
  "messages": [
    {
      "id": "1357917",
      "postDate": "06/20/2021 04:20:03",
      "content": "<p>when i use yolo and run test.py ,</p>\n<pre><code>python test.py --img 512 --batch 32 --conf 0.2 --iou 0.1:mAP@0.5=0.918\npython test.py --img 512 --batch 32 --conf 0.15 --iou 0.1:mAP@0.5=0.824\npython test.py --img 512 --batch 32 --conf 0.1 --iou 0.1:mAP@0.5=0.750\npython test.py --img 512 --batch 32 --conf 0.01 --iou 0.1:mAP@0.5=0.589\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.1:mAP@0.5=0.526\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.3:mAP@0.5=0.499\npython test.py --img 512 --batch 32 ---conf 0.001 --iou 0.65:mAP@0.5=0.471\n</code></pre>\n<p>I don't know why mAP is so sensitive to these param, because i 'm a new to object detection.</p>\n<pre><code>names:\n- 0. opacity\nnc: 1\ntrain: ~/512img600study/fold0_tra.txt\nval: ~/512img600study/fold0_val.txt\n</code></pre>\n<p>i use these for --data</p>",
      "rawMarkdown": "when i use yolo and run test.py ,\n\n```\npython test.py --img 512 --batch 32 --conf 0.2 --iou 0.1:mAP@0.5=0.918\npython test.py --img 512 --batch 32 --conf 0.15 --iou 0.1:mAP@0.5=0.824\npython test.py --img 512 --batch 32 --conf 0.1 --iou 0.1:mAP@0.5=0.750\npython test.py --img 512 --batch 32 --conf 0.01 --iou 0.1:mAP@0.5=0.589\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.1:mAP@0.5=0.526\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.3:mAP@0.5=0.499\npython test.py --img 512 --batch 32 ---conf 0.001 --iou 0.65:mAP@0.5=0.471\n```\n\nI don't know why mAP is so sensitive to these param, because i 'm a new to object detection.\n\n\n```\nnames:\n- 0. opacity\nnc: 1\ntrain: ~/512img600study/fold0_tra.txt\nval: ~/512img600study/fold0_val.txt\n```\ni use these for --data",
      "votes": null
    },
    {
      "id": "1357971",
      "postDate": "06/20/2021 05:26:20",
      "content": "<p>It's all about non maximum suppression. The raw prediction may contain a huge number of overlapping or/and low confidence boxes, because it is asked to generate hundreds of detections by default. This will lower the precision in the mAP calculation. </p>\n<p>Yolo code: <a href=\"https://github.com/ultralytics/yolov5/blob/master/test.py\" target=\"_blank\">https://github.com/ultralytics/yolov5/blob/master/test.py</a><br>\nSomething about NMS: <a href=\"https://www.analyticsvidhya.com/blog/2020/08/selecting-the-right-bounding-box-using-non-max-suppression-with-implementation\" target=\"_blank\">https://www.analyticsvidhya.com/blog/2020/08/selecting-the-right-bounding-box-using-non-max-suppression-with-implementation</a></p>",
      "rawMarkdown": "It's all about non maximum suppression. The raw prediction may contain a huge number of overlapping or/and low confidence boxes, because it is asked to generate hundreds of detections by default. This will lower the precision in the mAP calculation. \n\nYolo code: https://github.com/ultralytics/yolov5/blob/master/test.py\nSomething about NMS: https://www.analyticsvidhya.com/blog/2020/08/selecting-the-right-bounding-box-using-non-max-suppression-with-implementation",
      "votes": null
    },
    {
      "id": "1357986",
      "postDate": "06/20/2021 05:43:52",
      "content": "<p>thanks for replying.</p>\n<p>so, when there are many false positives for the same object, the lower --iou &amp;the higher --conf , I can get high mAP but it takes long time, right?</p>",
      "rawMarkdown": "thanks for replying.\n\nso, when there are many false positives for the same object, the lower --iou &the higher --conf , I can get high mAP but it takes long time, right?",
      "votes": null
    },
    {
      "id": "1357989",
      "postDate": "06/20/2021 05:53:10",
      "content": "<p>I also use yolov5 with <a href=\"https://www.kaggle.com/h053473666/siim-cov19-yolov5-train\" target=\"_blank\">this param</a><br>\nbut mAP is around 0.5 when i change --iou &amp; --conf.<br>\nIt ’s the height of my confusion.</p>",
      "rawMarkdown": "I also use yolov5 with [this param](https://www.kaggle.com/h053473666/siim-cov19-yolov5-train)\nbut mAP is around 0.5 when i change --iou & --conf.\nIt ’s the height of my confusion.",
      "votes": null
    },
    {
      "id": "1358051",
      "postDate": "06/20/2021 06:49:39",
      "content": "<p>you can refer to: <a href=\"https://www.kaggle.com/its7171/map-understanding-with-code-and-its-tips\" target=\"_blank\">https://www.kaggle.com/its7171/map-understanding-with-code-and-its-tips</a></p>",
      "rawMarkdown": "you can refer to: https://www.kaggle.com/its7171/map-understanding-with-code-and-its-tips",
      "votes": null
    },
    {
      "id": "1358068",
      "postDate": "06/20/2021 07:06:13",
      "content": "<p><a href=\"https://www.kaggle.com/underwearfitting\" target=\"_blank\">@underwearfitting</a> Thank you very much!!</p>",
      "rawMarkdown": "underwearfitting Thank you very much!!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1357971,
      "author_name": "houndcl",
      "author_url": "",
      "post_date": "06/20/2021 05:26:20",
      "content": "<p>It's all about non maximum suppression. The raw prediction may contain a huge number of overlapping or/and low confidence boxes, because it is asked to generate hundreds of detections by default. This will lower the precision in the mAP calculation. </p>\n<p>Yolo code: <a href=\"https://github.com/ultralytics/yolov5/blob/master/test.py\" target=\"_blank\">https://github.com/ultralytics/yolov5/blob/master/test.py</a><br>\nSomething about NMS: <a href=\"https://www.analyticsvidhya.com/blog/2020/08/selecting-the-right-bounding-box-using-non-max-suppression-with-implementation\" target=\"_blank\">https://www.analyticsvidhya.com/blog/2020/08/selecting-the-right-bounding-box-using-non-max-suppression-with-implementation</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1357986,
          "author_name": "abebe9849",
          "author_url": "",
          "post_date": "06/20/2021 05:43:52",
          "content": "<p>thanks for replying.</p>\n<p>so, when there are many false positives for the same object, the lower --iou &amp;the higher --conf , I can get high mAP but it takes long time, right?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1358051,
          "author_name": "underwearfitting",
          "author_url": "",
          "post_date": "06/20/2021 06:49:39",
          "content": "<p>you can refer to: <a href=\"https://www.kaggle.com/its7171/map-understanding-with-code-and-its-tips\" target=\"_blank\">https://www.kaggle.com/its7171/map-understanding-with-code-and-its-tips</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1358068,
          "author_name": "abebe9849",
          "author_url": "",
          "post_date": "06/20/2021 07:06:13",
          "content": "<p><a href=\"https://www.kaggle.com/underwearfitting\" target=\"_blank\">@underwearfitting</a> Thank you very much!!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1357989,
      "author_name": "abebe9849",
      "author_url": "",
      "post_date": "06/20/2021 05:53:10",
      "content": "<p>I also use yolov5 with <a href=\"https://www.kaggle.com/h053473666/siim-cov19-yolov5-train\" target=\"_blank\">this param</a><br>\nbut mAP is around 0.5 when i change --iou &amp; --conf.<br>\nIt ’s the height of my confusion.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1357917": "when i use yolo and run test.py ,\n\n```\npython test.py --img 512 --batch 32 --conf 0.2 --iou 0.1:mAP@0.5=0.918\npython test.py --img 512 --batch 32 --conf 0.15 --iou 0.1:mAP@0.5=0.824\npython test.py --img 512 --batch 32 --conf 0.1 --iou 0.1:mAP@0.5=0.750\npython test.py --img 512 --batch 32 --conf 0.01 --iou 0.1:mAP@0.5=0.589\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.1:mAP@0.5=0.526\npython test.py --img 512 --batch 32 --conf 0.001 --iou 0.3:mAP@0.5=0.499\npython test.py --img 512 --batch 32 ---conf 0.001 --iou 0.65:mAP@0.5=0.471\n```\n\nI don't know why mAP is so sensitive to these param, because i 'm a new to object detection.\n\n\n```\nnames:\n- 0. opacity\nnc: 1\ntrain: ~/512img600study/fold0_tra.txt\nval: ~/512img600study/fold0_val.txt\n```\ni use these for --data",
    "1357971": "It's all about non maximum suppression. The raw prediction may contain a huge number of overlapping or/and low confidence boxes, because it is asked to generate hundreds of detections by default. This will lower the precision in the mAP calculation. \n\nYolo code: https://github.com/ultralytics/yolov5/blob/master/test.py\nSomething about NMS: https://www.analyticsvidhya.com/blog/2020/08/selecting-the-right-bounding-box-using-non-max-suppression-with-implementation",
    "1357986": "thanks for replying.\n\nso, when there are many false positives for the same object, the lower --iou &the higher --conf , I can get high mAP but it takes long time, right?",
    "1357989": "I also use yolov5 with [this param](https://www.kaggle.com/h053473666/siim-cov19-yolov5-train)\nbut mAP is around 0.5 when i change --iou & --conf.\nIt ’s the height of my confusion.",
    "1358051": "you can refer to: https://www.kaggle.com/its7171/map-understanding-with-code-and-its-tips",
    "1358068": "underwearfitting Thank you very much!!"
  },
  "source": "meta"
}