{
  "id": 106607,
  "title": "Negative labels ",
  "url": "/competitions/open-images-2019-instance-segmentation/discussion/106607",
  "author_name": "",
  "post_date": "2019-08-30T08:15:26.857301100Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>It maybe a simple question but all info is scattered across multiple docs... so it is hard to find. Do we have separate training/validation files with 'negative' labels? I do not see any in the challenge-2019-validation-segmentation-masks.csv or challenge-2019-train-segmentation-masks.csv</p>",
  "messages": [
    {
      "id": "613108",
      "postDate": "08/30/2019 08:15:26",
      "content": "<p>It maybe a simple question but all info is scattered across multiple docs... so it is hard to find. Do we have separate training/validation files with 'negative' labels? I do not see any in the challenge-2019-validation-segmentation-masks.csv or challenge-2019-train-segmentation-masks.csv</p>",
      "rawMarkdown": "It maybe a simple question but all info is scattered across multiple docs... so it is hard to find. Do we have separate training/validation files with 'negative' labels? I do not see any in the challenge-2019-validation-segmentation-masks.csv or challenge-2019-train-segmentation-masks.csv",
      "votes": null
    },
    {
      "id": "613278",
      "postDate": "08/30/2019 11:40:02",
      "content": "<p>In the <a href=\"https://storage.googleapis.com/openimages/web/challenge2019_downloads.html\">challenge download page</a> you will find \"Image-level labels\" files (per track and add set).</p>\n\n<p>The data format itself is described in the <a href=\"https://storage.googleapis.com/openimages/web/download.html\">main download page</a>.</p>\n\n<blockquote>\n  <p>Confidence: Labels that are human-verified to be present in an image have confidence = 1 (positive labels). Labels that are human-verified to be absent from an image have confidence = 0 (negative labels). </p>\n</blockquote>\n\n<p>I hope this helps.\nLooking forward to see how participants use the negative labels!</p>",
      "rawMarkdown": "In the [challenge download page](https://storage.googleapis.com/openimages/web/challenge2019_downloads.html) you will find \"Image-level labels\" files (per track and add set).\n\nThe data format itself is described in the [main download page](https://storage.googleapis.com/openimages/web/download.html).\n\n&gt; Confidence: Labels that are human-verified to be present in an image have confidence = 1 (positive labels). Labels that are human-verified to be absent from an image have confidence = 0 (negative labels). \n\nI hope this helps.\nLooking forward to see how participants use the negative labels!",
      "votes": null
    },
    {
      "id": "614090",
      "postDate": "08/31/2019 04:13:31",
      "content": "<p>Thank you, it makes sense now. So, to clarify:\nfrom \"challenge-2019-validation-segmentation-labels.csv\"\nImageID,LabelName,Confidence\nd6d443cf4233a5b4,/m/03bt1vf,0\ne6fc75abc46fccc8,/m/0cmf2,1</p>\n\n<p>\"d6d443cf4233a5b4\" image does <strong>not</strong> have any \"/m/03bt1vf\".\n\"e6fc75abc46fccc8\" definitely has \"/m/0cmf2\", and ALL are segmented? I recall seeing somewhere that in some cases not all instances are segmented.</p>\n\n<p>Also, if \"d6d443cf4233a5b4\" only has that one entry '0' once, it means all other classes are undefined? i.e. '255'-not scored VOC-style?</p>",
      "rawMarkdown": "Thank you, it makes sense now. So, to clarify:\nfrom \"challenge-2019-validation-segmentation-labels.csv\"\nImageID,LabelName,Confidence\nd6d443cf4233a5b4,/m/03bt1vf,0\ne6fc75abc46fccc8,/m/0cmf2,1\n\n\"d6d443cf4233a5b4\" image does **not** have any \"/m/03bt1vf\".\n\"e6fc75abc46fccc8\" definitely has \"/m/0cmf2\", and ALL are segmented? I recall seeing somewhere that in some cases not all instances are segmented.\n\nAlso, if \"d6d443cf4233a5b4\" only has that one entry '0' once, it means all other classes are undefined? i.e. '255'-not scored VOC-style?",
      "votes": null
    },
    {
      "id": "614429",
      "postDate": "08/31/2019 13:23:25",
      "content": "<blockquote>\n  <p>I recall seeing somewhere that in some cases not all instances are segmented.</p>\n</blockquote>\n\n<p>Please read <a href=\"https://storage.googleapis.com/openimages/web/factsfigures.html\">description page</a>, the paragraph starting with \"We annotated all boxed instances of these 350 classes on the training split that fulfill the following criteria\".</p>\n\n<blockquote>\n  <p>It means all other classes are undefined?</p>\n</blockquote>\n\n<p>At the image level annotations, yes. That means that human raters gave no opinion regarding the other classes.</p>\n\n<p>You can get some extra information by using the machine generated image labels available at the <a href=\"https://storage.googleapis.com/openimages/web/download.html\">oidv5 download page</a>. These will be denser (regarding number of labels per image), but might have more errors than human-verified labels.</p>",
      "rawMarkdown": "&gt;  I recall seeing somewhere that in some cases not all instances are segmented.\n\nPlease read [description page](https://storage.googleapis.com/openimages/web/factsfigures.html), the paragraph starting with \"We annotated all boxed instances of these 350 classes on the training split that fulfill the following criteria\".\n\n&gt; It means all other classes are undefined?\n\nAt the image level annotations, yes. That means that human raters gave no opinion regarding the other classes.\n\nYou can get some extra information by using the machine generated image labels available at the [oidv5 download page](https://storage.googleapis.com/openimages/web/download.html). These will be denser (regarding number of labels per image), but might have more errors than human-verified labels.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 613278,
      "author_name": "benenson",
      "author_url": "",
      "post_date": "08/30/2019 11:40:02",
      "content": "<p>In the <a href=\"https://storage.googleapis.com/openimages/web/challenge2019_downloads.html\">challenge download page</a> you will find \"Image-level labels\" files (per track and add set).</p>\n\n<p>The data format itself is described in the <a href=\"https://storage.googleapis.com/openimages/web/download.html\">main download page</a>.</p>\n\n<blockquote>\n  <p>Confidence: Labels that are human-verified to be present in an image have confidence = 1 (positive labels). Labels that are human-verified to be absent from an image have confidence = 0 (negative labels). </p>\n</blockquote>\n\n<p>I hope this helps.\nLooking forward to see how participants use the negative labels!</p>",
      "votes": null,
      "replies": [
        {
          "id": 614090,
          "author_name": "dmitrykonovalov",
          "author_url": "",
          "post_date": "08/31/2019 04:13:31",
          "content": "<p>Thank you, it makes sense now. So, to clarify:\nfrom \"challenge-2019-validation-segmentation-labels.csv\"\nImageID,LabelName,Confidence\nd6d443cf4233a5b4,/m/03bt1vf,0\ne6fc75abc46fccc8,/m/0cmf2,1</p>\n\n<p>\"d6d443cf4233a5b4\" image does <strong>not</strong> have any \"/m/03bt1vf\".\n\"e6fc75abc46fccc8\" definitely has \"/m/0cmf2\", and ALL are segmented? I recall seeing somewhere that in some cases not all instances are segmented.</p>\n\n<p>Also, if \"d6d443cf4233a5b4\" only has that one entry '0' once, it means all other classes are undefined? i.e. '255'-not scored VOC-style?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 614429,
          "author_name": "benenson",
          "author_url": "",
          "post_date": "08/31/2019 13:23:25",
          "content": "<blockquote>\n  <p>I recall seeing somewhere that in some cases not all instances are segmented.</p>\n</blockquote>\n\n<p>Please read <a href=\"https://storage.googleapis.com/openimages/web/factsfigures.html\">description page</a>, the paragraph starting with \"We annotated all boxed instances of these 350 classes on the training split that fulfill the following criteria\".</p>\n\n<blockquote>\n  <p>It means all other classes are undefined?</p>\n</blockquote>\n\n<p>At the image level annotations, yes. That means that human raters gave no opinion regarding the other classes.</p>\n\n<p>You can get some extra information by using the machine generated image labels available at the <a href=\"https://storage.googleapis.com/openimages/web/download.html\">oidv5 download page</a>. These will be denser (regarding number of labels per image), but might have more errors than human-verified labels.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "613108": "It maybe a simple question but all info is scattered across multiple docs... so it is hard to find. Do we have separate training/validation files with 'negative' labels? I do not see any in the challenge-2019-validation-segmentation-masks.csv or challenge-2019-train-segmentation-masks.csv",
    "613278": "In the [challenge download page](https://storage.googleapis.com/openimages/web/challenge2019_downloads.html) you will find \"Image-level labels\" files (per track and add set).\n\nThe data format itself is described in the [main download page](https://storage.googleapis.com/openimages/web/download.html).\n\n&gt; Confidence: Labels that are human-verified to be present in an image have confidence = 1 (positive labels). Labels that are human-verified to be absent from an image have confidence = 0 (negative labels). \n\nI hope this helps.\nLooking forward to see how participants use the negative labels!",
    "614090": "Thank you, it makes sense now. So, to clarify:\nfrom \"challenge-2019-validation-segmentation-labels.csv\"\nImageID,LabelName,Confidence\nd6d443cf4233a5b4,/m/03bt1vf,0\ne6fc75abc46fccc8,/m/0cmf2,1\n\n\"d6d443cf4233a5b4\" image does **not** have any \"/m/03bt1vf\".\n\"e6fc75abc46fccc8\" definitely has \"/m/0cmf2\", and ALL are segmented? I recall seeing somewhere that in some cases not all instances are segmented.\n\nAlso, if \"d6d443cf4233a5b4\" only has that one entry '0' once, it means all other classes are undefined? i.e. '255'-not scored VOC-style?",
    "614429": "&gt;  I recall seeing somewhere that in some cases not all instances are segmented.\n\nPlease read [description page](https://storage.googleapis.com/openimages/web/factsfigures.html), the paragraph starting with \"We annotated all boxed instances of these 350 classes on the training split that fulfill the following criteria\".\n\n&gt; It means all other classes are undefined?\n\nAt the image level annotations, yes. That means that human raters gave no opinion regarding the other classes.\n\nYou can get some extra information by using the machine generated image labels available at the [oidv5 download page](https://storage.googleapis.com/openimages/web/download.html). These will be denser (regarding number of labels per image), but might have more errors than human-verified labels."
  },
  "source": "meta"
}