{
  "id": 61206,
  "title": "Can we talk about the class prediction confidence",
  "url": "/competitions/google-ai-open-images-object-detection-track/discussion/61206",
  "author_name": "Adams",
  "post_date": "2018-07-15T20:26:43.623000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi guys. From what I understand, for each bbox we have a confidence that a given image piece belongs to a class. Can someone tell me how should we treat confidences like 0.5 or 0?</p>\n\n<p>I understand that we should force the net to produce 0 for such an image, but what about the test set? Does it not matter what class we predict until we generate 0 for the specified class? Anyone wrapped their mind about this confidence stuff?</p>",
  "messages": [
    {
      "id": 357369,
      "postDate": "2018-07-15T20:26:43.623Z",
      "content": "<p>Hi guys. From what I understand, for each bbox we have a confidence that a given image piece belongs to a class. Can someone tell me how should we treat confidences like 0.5 or 0?</p>\n\n<p>I understand that we should force the net to produce 0 for such an image, but what about the test set? Does it not matter what class we predict until we generate 0 for the specified class? Anyone wrapped their mind about this confidence stuff?</p>",
      "rawMarkdown": "Hi guys. From what I understand, for each bbox we have a confidence that a given image piece belongs to a class. Can someone tell me how should we treat confidences like 0.5 or 0?\n\nI understand that we should force the net to produce 0 for such an image, but what about the test set? Does it not matter what class we predict until we generate 0 for the specified class? Anyone wrapped their mind about this confidence stuff?\n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "357369": "Hi guys. From what I understand, for each bbox we have a confidence that a given image piece belongs to a class. Can someone tell me how should we treat confidences like 0.5 or 0?\n\nI understand that we should force the net to produce 0 for such an image, but what about the test set? Does it not matter what class we predict until we generate 0 for the specified class? Anyone wrapped their mind about this confidence stuff?\n"
  }
}