{
  "id": 45814,
  "title": "were labels collected during, or after, the time when scans were made?",
  "url": "/competitions/passenger-screening-algorithm-challenge/discussion/45814",
  "author_name": "",
  "post_date": "2017-12-16T03:41:35.931092500Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I could imagine at least two methods to generate labels: i) a subject went to scan, and then somebody took the note on site; this would be the most accurate way of labelling the data. ii) afterwards, human raters label the data by looking at the scan images. </p>\n\n<p>I suspect that the labels were generated by using method ii) and raters looked at APS images.  My theory is that, APS data should be more lossy than A3DAPS data. However, it seems that models based on APS data performed better than A3DAPS data. See discussion here <a href=\"https://www.kaggle.com/c/passenger-screening-algorithm-challenge/discussion/45795\">https://www.kaggle.com/c/passenger-screening-algorithm-challenge/discussion/45795</a></p>",
  "messages": [
    {
      "id": "258418",
      "postDate": "12/16/2017 03:41:35",
      "content": "<p>I could imagine at least two methods to generate labels: i) a subject went to scan, and then somebody took the note on site; this would be the most accurate way of labelling the data. ii) afterwards, human raters label the data by looking at the scan images. </p>\n\n<p>I suspect that the labels were generated by using method ii) and raters looked at APS images.  My theory is that, APS data should be more lossy than A3DAPS data. However, it seems that models based on APS data performed better than A3DAPS data. See discussion here <a href=\"https://www.kaggle.com/c/passenger-screening-algorithm-challenge/discussion/45795\">https://www.kaggle.com/c/passenger-screening-algorithm-challenge/discussion/45795</a></p>",
      "rawMarkdown": "I could imagine at least two methods to generate labels: i) a subject went to scan, and then somebody took the note on site; this would be the most accurate way of labelling the data. ii) afterwards, human raters label the data by looking at the scan images. \n\nI suspect that the labels were generated by using method ii) and raters looked at APS images.  My theory is that, APS data should be more lossy than A3DAPS data. However, it seems that models based on APS data performed better than A3DAPS data. See discussion here https://www.kaggle.com/c/passenger-screening-algorithm-challenge/discussion/45795",
      "votes": null
    },
    {
      "id": "258427",
      "postDate": "12/16/2017 04:02:26",
      "content": "<p>I think they took the labels on site. I hand segmented the training aps threats and noticed there were a few cases where the threat was basically impossible to see (many of these were zone 9 threats and I suspect the threat was stuck right in between their legs). I don't think aps images are any lossier than a3daps data though, it seemed like a lot of the threats (especially the ones on zone 2/4) were easier to see in aps.</p>",
      "rawMarkdown": "I think they took the labels on site. I hand segmented the training aps threats and noticed there were a few cases where the threat was basically impossible to see (many of these were zone 9 threats and I suspect the threat was stuck right in between their legs). I don't think aps images are any lossier than a3daps data though, it seemed like a lot of the threats (especially the ones on zone 2/4) were easier to see in aps.",
      "votes": null
    },
    {
      "id": "258429",
      "postDate": "12/16/2017 04:08:53",
      "content": "<p>Good to know, suchir!</p>",
      "rawMarkdown": "Good to know, suchir!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 258427,
      "author_name": "suchir",
      "author_url": "",
      "post_date": "12/16/2017 04:02:26",
      "content": "<p>I think they took the labels on site. I hand segmented the training aps threats and noticed there were a few cases where the threat was basically impossible to see (many of these were zone 9 threats and I suspect the threat was stuck right in between their legs). I don't think aps images are any lossier than a3daps data though, it seemed like a lot of the threats (especially the ones on zone 2/4) were easier to see in aps.</p>",
      "votes": null,
      "replies": [
        {
          "id": 258429,
          "author_name": "luckyguy",
          "author_url": "",
          "post_date": "12/16/2017 04:08:53",
          "content": "<p>Good to know, suchir!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "258418": "I could imagine at least two methods to generate labels: i) a subject went to scan, and then somebody took the note on site; this would be the most accurate way of labelling the data. ii) afterwards, human raters label the data by looking at the scan images. \n\nI suspect that the labels were generated by using method ii) and raters looked at APS images.  My theory is that, APS data should be more lossy than A3DAPS data. However, it seems that models based on APS data performed better than A3DAPS data. See discussion here https://www.kaggle.com/c/passenger-screening-algorithm-challenge/discussion/45795",
    "258427": "I think they took the labels on site. I hand segmented the training aps threats and noticed there were a few cases where the threat was basically impossible to see (many of these were zone 9 threats and I suspect the threat was stuck right in between their legs). I don't think aps images are any lossier than a3daps data though, it seemed like a lot of the threats (especially the ones on zone 2/4) were easier to see in aps.",
    "258429": "Good to know, suchir!"
  },
  "source": "meta"
}