{
  "id": 121829,
  "title": "Metric for repeated prediction?",
  "url": "/competitions/pku-autonomous-driving/discussion/121829",
  "author_name": "",
  "post_date": "2019-12-15T21:01:14.114562800Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi! I have question on metric.</p>\n\n<p>The metric says:\n&gt; We then take the resulting distances between all pairs of objects and determine which predicted objects are closest to solution objects, and apply thresholds for both translation and rotation. Confidence scores are used to sort submission objects. Units for rotation are radians; translation is meters.</p>\n\n<p>So does that mean that two bounding boxes on one object at the same location will not affect the  score?</p>\n\n<p>Like this one:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fce96a9224c092611387815fe6f3938eb%2FID_8edf7f1e2.jpg?generation=1576443729418564&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": "695918",
      "postDate": "12/15/2019 21:01:14",
      "content": "<p>Hi! I have question on metric.</p>\n\n<p>The metric says:\n&gt; We then take the resulting distances between all pairs of objects and determine which predicted objects are closest to solution objects, and apply thresholds for both translation and rotation. Confidence scores are used to sort submission objects. Units for rotation are radians; translation is meters.</p>\n\n<p>So does that mean that two bounding boxes on one object at the same location will not affect the  score?</p>\n\n<p>Like this one:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fce96a9224c092611387815fe6f3938eb%2FID_8edf7f1e2.jpg?generation=1576443729418564&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "Hi! I have question on metric.\n\nThe metric says:\n&gt; We then take the resulting distances between all pairs of objects and determine which predicted objects are closest to solution objects, and apply thresholds for both translation and rotation. Confidence scores are used to sort submission objects. Units for rotation are radians; translation is meters.\n\nSo does that mean that two bounding boxes on one object at the same location will not affect the  score?\n\nLike this one:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fce96a9224c092611387815fe6f3938eb%2FID_8edf7f1e2.jpg?generation=1576443729418564&amp;alt=media)",
      "votes": null
    },
    {
      "id": "696112",
      "postDate": "12/16/2019 06:09:54",
      "content": "<p>The better fit will be used as the TP, the other(s) will count as FN. That's how I am reading this. </p>",
      "rawMarkdown": "The better fit will be used as the TP, the other(s) will count as FN. That's how I am reading this.",
      "votes": null
    },
    {
      "id": "696125",
      "postDate": "12/16/2019 06:37:04",
      "content": "<p>I think the metric only calculates the TP and FP. So my opinion is that since both bounding boxes on the same car are similar and accurate enough to pass the threshold, it shouldn't be a problem that there are two boxes. Is my understanding correct?</p>",
      "rawMarkdown": "I think the metric only calculates the TP and FP. So my opinion is that since both bounding boxes on the same car are similar and accurate enough to pass the threshold, it shouldn't be a problem that there are two boxes. Is my understanding correct?",
      "votes": null
    },
    {
      "id": "696467",
      "postDate": "12/16/2019 16:31:15",
      "content": "<p>FN is included now</p>",
      "rawMarkdown": "FN is included now",
      "votes": null
    },
    {
      "id": "696558",
      "postDate": "12/16/2019 18:49:07",
      "content": "<p>I'll take another look. Thanks!</p>",
      "rawMarkdown": "I'll take another look. Thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 696112,
      "author_name": "ilu000",
      "author_url": "",
      "post_date": "12/16/2019 06:09:54",
      "content": "<p>The better fit will be used as the TP, the other(s) will count as FN. That's how I am reading this. </p>",
      "votes": null,
      "replies": [
        {
          "id": 696125,
          "author_name": "tonychenxyz",
          "author_url": "",
          "post_date": "12/16/2019 06:37:04",
          "content": "<p>I think the metric only calculates the TP and FP. So my opinion is that since both bounding boxes on the same car are similar and accurate enough to pass the threshold, it shouldn't be a problem that there are two boxes. Is my understanding correct?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 696467,
          "author_name": "ilu000",
          "author_url": "",
          "post_date": "12/16/2019 16:31:15",
          "content": "<p>FN is included now</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 696558,
          "author_name": "tonychenxyz",
          "author_url": "",
          "post_date": "12/16/2019 18:49:07",
          "content": "<p>I'll take another look. Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "695918": "Hi! I have question on metric.\n\nThe metric says:\n&gt; We then take the resulting distances between all pairs of objects and determine which predicted objects are closest to solution objects, and apply thresholds for both translation and rotation. Confidence scores are used to sort submission objects. Units for rotation are radians; translation is meters.\n\nSo does that mean that two bounding boxes on one object at the same location will not affect the  score?\n\nLike this one:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fce96a9224c092611387815fe6f3938eb%2FID_8edf7f1e2.jpg?generation=1576443729418564&amp;alt=media)",
    "696112": "The better fit will be used as the TP, the other(s) will count as FN. That's how I am reading this.",
    "696125": "I think the metric only calculates the TP and FP. So my opinion is that since both bounding boxes on the same car are similar and accurate enough to pass the threshold, it shouldn't be a problem that there are two boxes. Is my understanding correct?",
    "696467": "FN is included now",
    "696558": "I'll take another look. Thanks!"
  },
  "source": "meta"
}