{
  "id": 215290,
  "title": "What is the role of the confidence in calculating your score?",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/215290",
  "author_name": "Jeremy",
  "post_date": "2021-01-29T10:34:54.194000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>As I understand it, the mAP is the mean over all classes of their area under their precision recall curves. Which means you only need the bounding box coordinates to calculate the mAP, as the bounding boxes are used to calculate IoU which is used to calculate precision/recall.</p>\n<p>So as I understand it, the confidence score isn't really needed to calculate mAP, or am i missing something? </p>",
  "messages": [
    {
      "id": 1175826,
      "postDate": "2021-01-29T11:17:52.003Z",
      "content": "<p>It is needed in case of multiple detections. Predictions in Pascal VOC metric are first sorted by confidence, then evaluated as TP/FP. Thus if you have many boxes associated to a single ground-truth box, only the one with highest confidence is considered as TP and all the others are marked as FP.</p>",
      "rawMarkdown": "It is needed in case of multiple detections. Predictions in Pascal VOC metric are first sorted by confidence, then evaluated as TP/FP. Thus if you have many boxes associated to a single ground-truth box, only the one with highest confidence is considered as TP and all the others are marked as FP.",
      "votes": 2,
      "replies": [
        {
          "id": 1175928,
          "postDate": "2021-01-29T12:08:35.207Z",
          "content": "<p>thanks for answering! one more thing-</p>\n<p>if you have two boxes which are really close, then the VOC metric would classify the lower confidence one as FP which actually lowers your score. Is the purpose of the non max supression to remove that lower confidence box to actually increase your score?</p>",
          "rawMarkdown": "thanks for answering! one more thing-\n\nif you have two boxes which are really close, then the VOC metric would classify the lower confidence one as FP which actually lowers your score. Is the purpose of the non max supression to remove that lower confidence box to actually increase your score?"
        },
        {
          "id": 1175986,
          "postDate": "2021-01-29T12:41:36.123Z",
          "content": "<p>Yes, NMS suppress boxes with an IoU bigger than a given threshold with the purpose of removing potential multiple detections </p>",
          "rawMarkdown": "Yes, NMS suppress boxes with an IoU bigger than a given threshold with the purpose of removing potential multiple detections ",
          "votes": 1
        }
      ]
    },
    {
      "id": 1175764,
      "postDate": "2021-01-29T10:34:54.193Z",
      "content": "<p>As I understand it, the mAP is the mean over all classes of their area under their precision recall curves. Which means you only need the bounding box coordinates to calculate the mAP, as the bounding boxes are used to calculate IoU which is used to calculate precision/recall.</p>\n<p>So as I understand it, the confidence score isn't really needed to calculate mAP, or am i missing something? </p>",
      "rawMarkdown": "As I understand it, the mAP is the mean over all classes of their area under their precision recall curves. Which means you only need the bounding box coordinates to calculate the mAP, as the bounding boxes are used to calculate IoU which is used to calculate precision/recall.\n\nSo as I understand it, the confidence score isn't really needed to calculate mAP, or am i missing something? ",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 1175826,
      "author_name": "LAZCoder",
      "author_url": "",
      "post_date": "2021-01-29T11:17:52.003000",
      "content": "<p>It is needed in case of multiple detections. Predictions in Pascal VOC metric are first sorted by confidence, then evaluated as TP/FP. Thus if you have many boxes associated to a single ground-truth box, only the one with highest confidence is considered as TP and all the others are marked as FP.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1175928,
          "author_name": "Jeremy",
          "author_url": "",
          "post_date": "2021-01-29T12:08:35.207000",
          "content": "<p>thanks for answering! one more thing-</p>\n<p>if you have two boxes which are really close, then the VOC metric would classify the lower confidence one as FP which actually lowers your score. Is the purpose of the non max supression to remove that lower confidence box to actually increase your score?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1175986,
          "author_name": "LAZCoder",
          "author_url": "",
          "post_date": "2021-01-29T12:41:36.123000",
          "content": "<p>Yes, NMS suppress boxes with an IoU bigger than a given threshold with the purpose of removing potential multiple detections </p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1175826": "It is needed in case of multiple detections. Predictions in Pascal VOC metric are first sorted by confidence, then evaluated as TP/FP. Thus if you have many boxes associated to a single ground-truth box, only the one with highest confidence is considered as TP and all the others are marked as FP.",
    "1175764": "As I understand it, the mAP is the mean over all classes of their area under their precision recall curves. Which means you only need the bounding box coordinates to calculate the mAP, as the bounding boxes are used to calculate IoU which is used to calculate precision/recall.\n\nSo as I understand it, the confidence score isn't really needed to calculate mAP, or am i missing something? "
  }
}