{
  "id": 426067,
  "title": "What's the intuition behind this metric?",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/426067",
  "author_name": "Gunes Evitan",
  "post_date": "2023-07-21T14:11:38.275000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>You evaluate detection quality based on IoU thresholds normally but does it make sense to evaluate their quality based on confidence thresholds? I couldn't think of any practical usage for this metric.</p>\n<p>Let's say I keep every detection and sort them by their confidence. The main idea is having a model that can generate high confidence correct detections but doesn't COCO mAP also achieves that? What am I missing here? </p>",
  "messages": [
    {
      "id": 2353216,
      "postDate": "2023-07-21T14:11:38.277Z",
      "content": "<p>You evaluate detection quality based on IoU thresholds normally but does it make sense to evaluate their quality based on confidence thresholds? I couldn't think of any practical usage for this metric.</p>\n<p>Let's say I keep every detection and sort them by their confidence. The main idea is having a model that can generate high confidence correct detections but doesn't COCO mAP also achieves that? What am I missing here? </p>",
      "rawMarkdown": "You evaluate detection quality based on IoU thresholds normally but does it make sense to evaluate their quality based on confidence thresholds? I couldn't think of any practical usage for this metric.\n\nLet's say I keep every detection and sort them by their confidence. The main idea is having a model that can generate high confidence correct detections but doesn't COCO mAP also achieves that? What am I missing here? ",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2353216": "You evaluate detection quality based on IoU thresholds normally but does it make sense to evaluate their quality based on confidence thresholds? I couldn't think of any practical usage for this metric.\n\nLet's say I keep every detection and sort them by their confidence. The main idea is having a model that can generate high confidence correct detections but doesn't COCO mAP also achieves that? What am I missing here? "
  }
}