{
  "id": 181260,
  "title": "Most test images do not have landmarks according to provided GT metadata...?",
  "url": "/competitions/landmark-recognition-2020/discussion/181260",
  "author_name": "",
  "post_date": "2020-09-08T07:48:54.356086300Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I am new to this, so please bear with me…</p>\n<p>I was skimming through the gldv2 git for data, and looked at </p>\n<p><a href=\"https://s3.amazonaws.com/google-landmark/ground_truth/recognition_solution_v2.1.csv\" target=\"_blank\">https://s3.amazonaws.com/google-landmark/ground_truth/recognition_solution_v2.1.csv</a><br>\n(retrieved from: <a href=\"https://github.com/cvdfoundation/google-landmark\" target=\"_blank\">https://github.com/cvdfoundation/google-landmark</a>)</p>\n<p>which seems like the GT landmark assignment for each test image.</p>\n<p>and when you open the csv file you may realize that most images are not assigned with a GT landmark… Some open solutions provided by kaggle notebook seemes like to make predictions for every test image, but if most test images do not include landmarks, then how can the GAP score be 0.48 even after making predictions for every test images…?????</p>\n<p>Please enlighten me…</p>",
  "messages": [
    {
      "id": "1002529",
      "postDate": "09/08/2020 07:48:54",
      "content": "<p>I am new to this, so please bear with me…</p>\n<p>I was skimming through the gldv2 git for data, and looked at </p>\n<p><a href=\"https://s3.amazonaws.com/google-landmark/ground_truth/recognition_solution_v2.1.csv\" target=\"_blank\">https://s3.amazonaws.com/google-landmark/ground_truth/recognition_solution_v2.1.csv</a><br>\n(retrieved from: <a href=\"https://github.com/cvdfoundation/google-landmark\" target=\"_blank\">https://github.com/cvdfoundation/google-landmark</a>)</p>\n<p>which seems like the GT landmark assignment for each test image.</p>\n<p>and when you open the csv file you may realize that most images are not assigned with a GT landmark… Some open solutions provided by kaggle notebook seemes like to make predictions for every test image, but if most test images do not include landmarks, then how can the GAP score be 0.48 even after making predictions for every test images…?????</p>\n<p>Please enlighten me…</p>",
      "rawMarkdown": "I am new to this, so please bear with me...\n\nI was skimming through the gldv2 git for data, and looked at \n\nhttps://s3.amazonaws.com/google-landmark/ground_truth/recognition_solution_v2.1.csv\n(retrieved from: https://github.com/cvdfoundation/google-landmark)\n\nwhich seems like the GT landmark assignment for each test image.\n\nand when you open the csv file you may realize that most images are not assigned with a GT landmark... Some open solutions provided by kaggle notebook seemes like to make predictions for every test image, but if most test images do not include landmarks, then how can the GAP score be 0.48 even after making predictions for every test images...?????\n\nPlease enlighten me...",
      "votes": null
    },
    {
      "id": "1003417",
      "postDate": "09/09/2020 00:32:09",
      "content": "<ol>\n<li>I doubt that the distribution in the GLDv2-clean is the same as in the public GLDv2. With that being said, would love to hear from organizers on that.</li>\n<li>The GAP score can be high if you make predictions with lower confidence on distractors.</li>\n</ol>",
      "rawMarkdown": "1. I doubt that the distribution in the GLDv2-clean is the same as in the public GLDv2. With that being said, would love to hear from organizers on that.\n2. The GAP score can be high if you make predictions with lower confidence on distractors.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1003417,
      "author_name": "chankhavu",
      "author_url": "",
      "post_date": "09/09/2020 00:32:09",
      "content": "<ol>\n<li>I doubt that the distribution in the GLDv2-clean is the same as in the public GLDv2. With that being said, would love to hear from organizers on that.</li>\n<li>The GAP score can be high if you make predictions with lower confidence on distractors.</li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1002529": "I am new to this, so please bear with me...\n\nI was skimming through the gldv2 git for data, and looked at \n\nhttps://s3.amazonaws.com/google-landmark/ground_truth/recognition_solution_v2.1.csv\n(retrieved from: https://github.com/cvdfoundation/google-landmark)\n\nwhich seems like the GT landmark assignment for each test image.\n\nand when you open the csv file you may realize that most images are not assigned with a GT landmark... Some open solutions provided by kaggle notebook seemes like to make predictions for every test image, but if most test images do not include landmarks, then how can the GAP score be 0.48 even after making predictions for every test images...?????\n\nPlease enlighten me...",
    "1003417": "1. I doubt that the distribution in the GLDv2-clean is the same as in the public GLDv2. With that being said, would love to hear from organizers on that.\n2. The GAP score can be high if you make predictions with lower confidence on distractors."
  },
  "source": "meta"
}